Communication method and device and readable storage medium

Abstract

The invention relates to the technical field of communication, in particular to a communication method and device and a readable storage medium, and the method comprises the steps that a first communication device in a mobile communication network sends first information to a second communication device; a control node of a large model is deployed on the first communication device, a first sub-neural network model in a plurality of sub-neural network models of the large model is deployed on the second communication device, the first information is used for indicating a first operator, and the first operator is used for compressing service data transmitted by the second communication device to other communication devices; other communication devices are deployed with other sub-neural network models except the first sub-neural network model in the plurality of sub-neural network models. The embodiment of the invention can improve the performance of a large model.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a communication method, device and readable storage medium. Background Art

[0002] In the embryonic stage of artificial intelligence technology, the emergence of neural network models represented by convolutional neural networks (CNN) has enabled in-depth research and application in fields such as natural language generation and computer vision. These technologies have gradually been applied to fields such as intelligent customer service, intelligent recommendation, and autonomous driving, providing the public with more convenient and efficient services. Subsequently, the emergence of new neural network models such as generative adversarial networks (GAN) and Transformer has further developed and improved the algorithm structure and training technology of large models, and the scale of the models has gradually increased until the emergence of models with hundreds of billions of parameters (such as BERT and GPT), and the capabilities of large models have been unprecedentedly improved. The application scenarios of these large models are also more extensive, including natural language processing, computer vision, speech recognition and other fields. For example, ChatGPT can accurately recognize and understand users' voice and text information, and provide more intelligent services. This has also promoted the realization of intelligent inclusiveness, allowing more people to enjoy the convenience brought by intelligent services.

[0003] The emergence of big models is changing all industries, and mobile communication networks are no exception.

[0004] The 44th meeting of the International Telecommunication Union Radiocommunication Sector Working Group 5D (ITU-R WP5D) formally adopted the "Recommendation on the Framework and Overall Objectives of IMT for 2030 and Beyond". The recommendation identifies "intelligent inclusion" as one of the core scenarios of 6G networks and a key feature that distinguishes it from 5G networks. The recommendation believes that 6G networks will become AI-enabled infrastructure and support intelligent applications in the form of services. It can be seen that mobile communication networks will also participate in the industry of providing large model services, give full play to their own characteristics, and bring true intelligent inclusion to users.

[0005] At present, solutions for combining large models with mobile communication networks are still being explored. Summary of the invention

[0006] The embodiments of the present application provide a communication method, device and readable storage medium, which can improve the performance of large models.

[0007] The present application is introduced below from different aspects. It should be understood that the implementation methods and beneficial effects of the following different aspects can be referenced to each other.

[0008] In a first aspect, the present application provides a communication method, which is applied to a first communication device in a mobile communication network, wherein a control node of a large model is deployed on the first communication device, and the large model also includes multiple sub-neural network models, which are respectively deployed on different communication devices in the mobile communication network. The method comprises: sending first information to a second communication device in the mobile communication network; a first sub-neural network model of the multiple sub-neural network models is deployed on the second communication device, the first information is used to indicate a first operator, and the first operator is used to compress business data transmitted by the second communication device to other communication devices, and the other communication devices are deployed with other sub-neural network models of the multiple sub-neural network models except the first sub-neural network model.

[0009] In an embodiment of the present application, when multiple task execution nodes (TE, corresponding to sub-neural network models) of a large model are respectively deployed on different communication devices in a mobile communication network, the network side can configure a network affinity operator (i.e., the first operator) for compressing the business data to be transmitted between the multiple task execution nodes, thereby reducing the data volume of the business data that needs to be transmitted between different task execution nodes. Even when the interface bandwidth between different communication devices is limited, data transmission can be completed effectively and timely, thereby improving the performance of the large model.

[0010] In a possible implementation, the method further includes:

[0011] Sending second information to a third communication device in the mobile communication network; the third communication device is deployed with a second sub-neural network model among the multiple sub-neural network models, the second information is used to indicate a second operator, and the second operator is used to decompress the business data compressed by the first operator.

[0012] In an embodiment of the present application, while the network side configures an affinity operator for a task execution node of the large model (such as deployed on the second communication device), it can also correspondingly configure a sparse operator (i.e., the second operator) for another task execution node of the large model (such as deployed on the third communication device). That is, the network side can configure a pair of affinity operators and sparse operators, which are used by the second communication device and the third communication device respectively, so that the third communication device can use the sparse operator to decompress the business data compressed by the affinity operator, thereby improving the performance of the large model while reducing information loss.

[0013] In a possible implementation, the first operator and / or the second operator are determined based on the model type and the model splitting point of the large model, or the first operator and / or the third operator are determined based on a training algorithm.

[0014] In an embodiment of the present application, the affinity operator and the sparse operator can be determined based on the model type of the large model and the model splitting point, that is, based on what kind of large model is deployed and how to split the deployment; alternatively, the affinity operator and the sparse operator can also be determined after training by a training algorithm.

[0015] In a possible implementation, the method further includes:

[0016] Sending third information to the second communication device; the third information includes an initial operator, and the initial operator is used to compress the service data transmitted by the second communication device to the other communication device;

[0017] Sending fourth information to a third communication device in the mobile communication network; the fourth information includes a loss function, and the first operator is obtained based on training of the loss function and the initial operator.

[0018] In an embodiment of the present application, for a method that requires determining an affinity operator through a training algorithm, the network side can first send an initial operator to the second communication device, and at the same time send a training loss function to the third communication device. After repeatedly training and tuning the initial operator until convergence, a trained affinity operator (such as the first operator) can be obtained.

[0019] In a possible implementation, the method further includes:

[0020] Sending fifth information to the second communication device; the fifth information is used to indicate a third operator, the third operator is used to compress the service data transmitted by the second communication device to the other communication devices, and the compression rate of the third operator is different from the compression rate of the first operator.

[0021] In the embodiment of the present application, after the network side sends the affinity operator to the second communication device, it can also reconfigure a new affinity operator for the second communication device according to the actual network situation, so as to meet the actual network transmission requirements.

[0022] In a possible implementation, the method further includes:

[0023] Sending sixth information to a third communication device in the mobile communication network; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the service data compressed by the third operator.

[0024] In the embodiment of the present application, while the network side reconfigures the affinity operator, it may also correspondingly reconfigure a sparse operator to minimize information loss.

[0025] In a possible implementation, the method further includes:

[0026] Sending seventh information to the second communication device; the seventh information includes first parameter information of the large model, and the seventh information is used to instruct the second communication device to load the first sub-neural network model of the large model based on the first parameter information;

[0027] Sending an eighth information to a third communication device in the mobile communication network; the eighth information includes the second parameter information of the large model, and the eighth information is used to instruct the third communication device to load the second sub-neural network model of the large model based on the second parameter information.

[0028] In an embodiment of the present application, after the network side determines the deployment plan of the large model in the mobile communication network, it can send instruction information to each communication device that needs to deploy the sub-neural network model of the large model, instructing these communication devices to load the corresponding sub-neural network models respectively, so that these communication devices have the ability to process large model services.

[0029] In a second aspect, the present application provides a communication method, which is applied to a second communication device in a mobile communication network, wherein a first sub-neural network model in a large model is deployed on the second communication device, wherein the large model includes multiple sub-neural network models, and the multiple sub-neural network models are respectively deployed on different communication devices in the mobile communication network; the method includes:

[0030] receiving first information sent by a first communication device in the mobile communication network; the first information is used to indicate a first operator; the large model further includes a control node, and the control node is deployed on the first communication device;

[0031] Acquire first business data, and input the first business data into the first sub-neural network model for processing to obtain second business data;

[0032] Sending third business data to a third communication device in the mobile communication network; the second sub-neural network model in the large model is deployed on the third communication device, and the third business data is obtained by compressing the second business data based on the first operator, and the data volume of the third business data is smaller than the data volume of the second business data.

[0033] In an embodiment of the present application, when multiple task execution nodes (TE, corresponding to sub-neural network models) of the large model are respectively deployed on different communication devices of the mobile communication network, the network side can configure a network affinity operator (i.e., the first operator) to the second communication device, so that after the second communication device uses the sub-neural network model to process the business data, it first compresses the processing result (i.e., the second business data) through the affinity operator to reduce the data level of the business data to be transmitted, and then sends the compressed result (i.e., the third business data) to the third communication device. In this way, when the interface bandwidth between the second communication device and the third communication device is limited, data transmission can be completed effectively and timely, thereby improving the performance of the large model.

[0034] In a possible implementation, the first operator is determined based on a model type and a model splitting point of the large model, or the first operator is determined by the second communication device and the third communication device based on a training algorithm.

[0035] In a possible implementation, the method further includes:

[0036] Receive third information sent by the first communication device; the third information includes an initial operator, the third information is used to instruct the second communication device to train the initial operator, the initial operator is used to compress the service data transmitted by the second communication device to other communication devices, and the first operator is obtained based on the training of the initial operator.

[0037] In one possible implementation, the data volume of the third business data is smaller than the data volume of the second business data, including: the second business data includes information of M neurons, the third business data includes information of N neurons, the N neurons are some of the M neurons, and N is a positive integer less than M.

[0038] In an embodiment of the present application, the second communication device uses an affinity operator to compress business data, which can be achieved by reducing the number of neurons and changing the way information is distributed on the original neurons, thereby transmitting only information of some neurons and achieving effective compression of the data to be transmitted.

[0039] In a possible implementation, the method further includes:

[0040] Receive fifth information sent by the first communication device; the fifth information is used to indicate a third operator; the third operator is used to compress business data transmitted by the second communication device to other communication devices, the compression rate of the third operator is different from the compression rate of the first operator, and the change value of the data transmission rate between the second communication device and the third communication device is greater than or equal to a preset threshold.

[0041] In an embodiment of the present application, when the change value of the data transmission rate between the second communication device and the third communication device is greater than or equal to a preset threshold, the network side can reconfigure a new affinity operator for the second communication device to adapt to the new data transmission rate to ensure the performance of the large model.

[0042] In a possible implementation, the method further includes:

[0043] receiving seventh information sent by the first communication device; the seventh information includes first parameter information of the large model;

[0044] The first sub-neural network model of the large model is loaded based on the first parameter information.

[0045] In a possible implementation, the first business data is a business request sent by a terminal device, and a response to the first business data is obtained based on processing of the third business data by the second sub-neural network model.

[0046] In an embodiment of the present application, the business data (i.e., business request) of the large model is initiated by the terminal device, and the business request can be responded to after being processed by the first sub-neural network model and the second sub-neural network.

[0047] In a third aspect, the present application provides a communication method, which is applied to a third communication device in a mobile communication network, wherein a second sub-neural network model in the large model is deployed on the third communication device, and the large model includes multiple sub-neural network models, and the multiple sub-neural network models are respectively deployed on different communication devices in the mobile communication network; the method includes:

[0048] receiving second information sent by a first communication device in the mobile communication network; the second information is used to indicate a second operator; the large model further includes a control node, and the control node is deployed on the first communication device;

[0049] Receiving third service data sent by a second communication device in the mobile communication network; the first sub-neural network model in the large model is deployed on the second communication device;

[0050] The fourth business data is input into the second sub-neural network model for processing to obtain the fifth business data, and the fifth business data is sent; the fourth business data is obtained by decompressing the third business data based on the second operator; the data volume of the third business data is smaller than the data volume of the fourth business data.

[0051] In an embodiment of the present application, while the network side configures an affinity operator for a task execution node of the large model (such as deployed on the second communication device), it can also correspondingly configure a sparse operator (i.e., the second operator) for another task execution node of the large model (such as deployed on the third communication device). That is, the network side can configure a pair of affinity operators and sparse operators, which are used by the second communication device and the third communication device respectively, so that the third communication device can use the sparse operator to decompress the business data compressed by the affinity operator, thereby improving the performance of the large model while reducing information loss.

[0052] In a possible implementation, the second operator is determined based on a model type and a model splitting point of the large model, or the second operator is determined by the first communication device and the second communication device based on a training algorithm.

[0053] In a possible implementation, the method further includes:

[0054] Receive fourth information sent by the first communication device; the fourth information includes a loss function, the first operator of the second communication device is obtained based on the training of the loss function, and the first operator is used to compress the service data transmitted by the second communication device to other communication devices.

[0055] In a possible implementation, the data volume of the third business data is smaller than the data volume of the fourth business data, including: the third business data includes information of N neurons, the fourth business data includes information of K neurons, the N neurons are some of the K neurons, and N is a positive integer less than K.

[0056] In the embodiment of the present application, the third communication device uses a sparse operator to decompress the business data, which can be done by redistributing the received neuron information to more neurons, thereby reducing the information density that may be caused by compression and reducing information loss.

[0057] In a possible implementation, the method further includes:

[0058] Receive sixth information sent by the first communication device; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the business data compressed by the third operator.

[0059] In a possible implementation, the method further includes:

[0060] receiving eighth information sent by the first communication device; the eighth information including second parameter information of the large model;

[0061] The second sub-neural network model of the large model is loaded based on the second parameter information.

[0062] In a possible implementation manner, the sending the fifth service data includes:

[0063] The fifth service data is sent to the terminal device, where the fifth service data is a response to the service request initiated by the terminal device.

[0064] In an embodiment of the present application, the last task execution node of the large model is deployed on the third communication device. After the third communication device completes processing of the business data, the processing result can be returned to the communication device (such as a terminal device) that initiated the reasoning request as a response to the reasoning request.

[0065] In a fourth aspect, the present application provides a first communication device in a mobile communication network, wherein a control node of a large model is deployed on the first communication device, the large model further includes a plurality of sub-neural network models, and the plurality of sub-neural network models are respectively deployed on different communication devices in the mobile communication network, the first communication device comprising:

[0066] A sending unit, used for sending first information to a second communication device in the mobile communication network; a first sub-neural network model among the multiple sub-neural network models is deployed on the second communication device, the first information is used to indicate a first operator, and the first operator is used to compress business data transmitted by the second communication device to other communication devices, and the other communication devices are deployed with other sub-neural network models among the multiple sub-neural network models except the first sub-neural network model.

[0067] In a possible implementation manner, the sending unit is further configured to:

[0068] Sending second information to a third communication device in the mobile communication network; the third communication device is deployed with a second sub-neural network model among the multiple sub-neural network models, the second information is used to indicate a second operator, and the second operator is used to decompress the business data compressed by the first operator.

[0069] In a possible implementation, the first operator and / or the second operator are determined based on the model type and the model splitting point of the large model, or the first operator and / or the third operator are determined based on a training algorithm.

[0070] In a possible implementation manner, the sending unit is further configured to:

[0071] Sending third information to the second communication device; the third information includes an initial operator, and the initial operator is used to compress the service data transmitted by the second communication device to the other communication device;

[0072] Sending fourth information to a third communication device in the mobile communication network; the fourth information includes a loss function, and the first operator is obtained based on training of the loss function and the initial operator.

[0073] In a possible implementation manner, the sending unit is further configured to:

[0074] Sending fifth information to the second communication device; the fifth information is used to indicate a third operator, the third operator is used to compress the service data transmitted by the second communication device to the other communication devices, and the compression rate of the third operator is different from the compression rate of the first operator.

[0075] In a possible implementation manner, the sending unit is further configured to:

[0076] Sending sixth information to a third communication device in the mobile communication network; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the service data compressed by the third operator.

[0077] In a possible implementation manner, the sending unit is further configured to:

[0078] Sending seventh information to the second communication device; the seventh information includes first parameter information of the large model, and the seventh information is used to instruct the second communication device to load the first sub-neural network model of the large model based on the first parameter information;

[0079] Sending an eighth information to a third communication device in the mobile communication network; the eighth information includes the second parameter information of the large model, and the eighth information is used to instruct the third communication device to load the second sub-neural network model of the large model based on the second parameter information.

[0080] In a fifth aspect, the present application provides a second communication device in a mobile communication network, on which a first sub-neural network model in a large model is deployed, the large model includes multiple sub-neural network models, and the multiple sub-neural network models are respectively deployed on different communication devices in the mobile communication network; the second communication device includes:

[0081] A receiving unit, configured to receive first information sent by a first communication device in the mobile communication network; the first information is used to indicate a first operator; the large model further includes a control node, and the control node is deployed on the first communication device;

[0082] A processing unit, used for acquiring first business data, and inputting the first business data into the first sub-neural network model for processing to obtain second business data;

[0083] A sending unit is used to send third business data to a third communication device in the mobile communication network; the second sub-neural network model in the large model is deployed on the third communication device, and the third business data is obtained by compressing the second business data based on the first operator, and the data volume of the third business data is smaller than the data volume of the second business data.

[0084] In a possible implementation, the first operator is determined based on a model type and a model splitting point of the large model, or the first operator is determined by the second communication device and the third communication device based on a training algorithm.

[0085] In a possible implementation manner, the receiving unit is further configured to:

[0086] Receive third information sent by the first communication device; the third information includes an initial operator, the third information is used to instruct the second communication device to train the initial operator, the initial operator is used to compress the service data transmitted by the second communication device to other communication devices, and the first operator is obtained based on the training of the initial operator.

[0087] In one possible implementation, the data volume of the third business data is smaller than the data volume of the second business data, including: the second business data includes information of M neurons, the third business data includes information of N neurons, the N neurons are some of the M neurons, and N is a positive integer less than M.

[0088] In a possible implementation manner, the receiving unit is further configured to:

[0089] Receive fifth information sent by the first communication device; the fifth information is used to indicate a third operator; the third operator is used to compress business data transmitted by the second communication device to other communication devices, the compression rate of the third operator is different from the compression rate of the first operator, and the change value of the data transmission rate between the second communication device and the third communication device is greater than or equal to a preset threshold.

[0090] In a possible implementation manner, the receiving unit is further configured to:

[0091] receiving seventh information sent by the first communication device; the seventh information includes first parameter information of the large model;

[0092] The first sub-neural network model of the large model is loaded based on the first parameter information.

[0093] In a possible implementation, the first business data is a business request sent by a terminal device, and a response to the first business data is obtained based on processing of the third business data by the second sub-neural network model.

[0094] In a sixth aspect, the present application provides a third communication device in a mobile communication network, on which a second sub-neural network model in the large model is deployed, the large model includes multiple sub-neural network models, and the multiple sub-neural network models are respectively deployed on different communication devices in the mobile communication network; the third communication device includes:

[0095] A receiving unit, configured to receive second information sent by a first communication device in the mobile communication network; the second information is used to indicate a second operator; the large model further includes a control node, and the control node is deployed on the first communication device;

[0096] The receiving unit is further used to receive third service data sent by a second communication device in the mobile communication network; the first sub-neural network model in the large model is deployed on the second communication device;

[0097] A processing unit, used for inputting the fourth business data into the second sub-neural network model for processing to obtain fifth business data;

[0098] A sending unit is used to send the fifth business data; the fourth business data is obtained by decompressing the third business data based on the second operator; the data volume of the third business data is smaller than the data volume of the fourth business data.

[0099] In a possible implementation, the second operator is determined based on a model type and a model splitting point of the large model, or the second operator is determined by the first communication device and the second communication device based on a training algorithm.

[0100] In a possible implementation manner, the receiving unit is further configured to:

[0101] Receive fourth information sent by the first communication device; the fourth information includes a loss function, the first operator of the second communication device is obtained based on the training of the loss function, and the first operator is used to compress the service data transmitted by the second communication device to other communication devices.

[0102] In a possible implementation, the data volume of the third business data is smaller than the data volume of the fourth business data, including: the third business data includes information of N neurons, the fourth business data includes information of K neurons, the N neurons are some of the K neurons, and N is a positive integer less than K.

[0103] In a possible implementation manner, the receiving unit is further configured to:

[0104] Receive sixth information sent by the first communication device; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the business data compressed by the third operator.

[0105] In a possible implementation manner, the receiving unit is further configured to:

[0106] receiving eighth information sent by the first communication device; the eighth information including second parameter information of the large model;

[0107] The second sub-neural network model of the large model is loaded based on the second parameter information.

[0108] In a possible implementation manner, the sending unit is specifically configured to:

[0109] The fifth service data is sent to the terminal device, where the fifth service data is a response to the service request initiated by the terminal device.

[0110] In a seventh aspect, the present application provides a communication device, which may include a processor, a transceiver, and a memory. The memory is used to store a computer program, the transceiver is used to send and receive various messages, and the computer program includes program instructions. When the processor runs the program instructions, the communication device executes the method described in any one of the first to third aspects, or any possible implementation of any one of the aspects. The transceiver may be a radio frequency module in the communication device, or a combination of a radio frequency module and an antenna, or an input and output interface of a chip or circuit.

[0111] In an eighth aspect, the present application provides a computer-readable storage medium having program instructions stored thereon, which, when executed on a computer, enables the computer to execute the method described in any one of the first to third aspects above, or any possible implementation of any one of the aspects.

[0112] In a ninth aspect, the present application provides a program product comprising program instructions, which, when executed, enables the method described in any one of the first to third aspects, or any possible implementation of any one of the aspects to be executed.

[0113] In the tenth aspect, the present application provides a communication device, which can be implemented in the form of a chip, or in the form of a device or a component in a device, etc. The device includes a processor. The processor is used to read and execute a program stored in a memory to execute the communication method provided by one or more of any possible implementations of any aspect from the first aspect to the third aspect, or any aspect thereof. Optionally, the communication device also includes a memory, which is connected to the processor through a circuit. Further optionally, the communication device also includes a communication interface, and the processor is connected to the communication interface. The communication interface is used to receive a data packet and / or information to be processed, and the processor obtains the data packet and / or information from the communication interface, processes the data packet and / or information, and outputs the processing result through the communication interface. The communication interface can be an input and output interface.

[0114] Optionally, the processor and memory may be physically independent units, or the memory may be integrated with the processor.

[0115] In an eleventh aspect, the present application provides a mobile communication system, which includes a first communication device, a second communication device and a third communication device; the first communication device is used to execute the method described in the first aspect above, or any possible implementation of any aspect thereof, the second communication device is used to execute the method described in the second aspect above, or any possible implementation of any aspect thereof, and the second communication device is used to execute the method described in the third aspect above, or any possible implementation of any aspect thereof.

[0116] The technical effects achieved in the above-mentioned aspects can be referred to each other or to the beneficial effects in the method embodiments shown below, and will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the background technology, the drawings required for use in the embodiments of the present application or the background technology will be described below.

[0118] Figure 1 It is a schematic diagram of a scenario architecture to which a communication method provided in an embodiment of the present application can be applied.

[0119] Figure 2a It is a schematic diagram of the architecture of a communication system provided in an embodiment of the present application.

[0120] Figure 2b It is a schematic diagram of the Sidelink UE-to-Network Relay scenario provided in an embodiment of the present application.

[0121] Figure 2c It is a schematic diagram of the Sidelink UE-to-UE Relay scenario provided in an embodiment of the present application.

[0122] Figure 3 It is a simplified structural diagram of a UE and a base station provided in an embodiment of the present application.

[0123] Figure 4 It is a structural diagram of a large model combined with a mobile communication network.

[0124] Figure 5a It is a flow chart of a communication method provided in an embodiment of the present application.

[0125] Figure 5b It is a schematic diagram of an affinity operator compression principle provided in an embodiment of the present application.

[0126] Figure 5c It is a structural schematic diagram of a large model splitting provided in an embodiment of the present application.

[0127] Figure 6a It is a flowchart of an affinity operator training method provided in an embodiment of the present application.

[0128] Figure 6b It is a data processing flow diagram of affinity operator training provided in an embodiment of the present application.

[0129] Figure 7a It is a flow chart of another communication method provided in an embodiment of the present application.

[0130] Figure 7b It is a schematic diagram of a sparse operator decompression principle provided in an embodiment of the present application.

[0131] Figure 8a It is a flowchart of an operator pair training method provided in an embodiment of the present application.

[0132] Figure 8b It is a schematic diagram of the data processing flow of an operator pair training provided in an embodiment of the present application.

[0133] Fig. 9 It is a structural diagram of a communication device provided in an embodiment of the present application.

[0134] Fig.10 It is a structural diagram of another communication device provided in an embodiment of the present application.

[0135] Fig.11 It is a structural diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0136] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0137] In the description of the present application, the words "first", "second", etc. are only used to distinguish different objects, and do not limit the quantity and execution order, and the words "first", "second", etc. are not necessarily different. For example, the first message and the second message are only used to distinguish different information, and their order is not limited. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0138] In the description of this application, unless otherwise specified, " / " means "or", for example, A / B can mean A or B. "And / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, "at least one (item)", "the following one (item) or more (items)" or similar expressions refer to any combination of these items, including any combination of single items (items) or plural items (items). For example, at least one item (item) of a, b, or c can mean: a, b, c; a and b; a and c; b and c; or a, b and c. Among them, a, b, c can be single or multiple.

[0139] In the description of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described in this application as "exemplary", "for example" or "for example" should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a concrete way.

[0140] It can be understood that in the description of this application, "when", "if" and "if" all mean that the device will take corresponding actions under certain objective circumstances, but do not limit the time, nor do they require that the device must have a judgment action when it is implemented, nor do they mean that there are other limitations.

[0141] The term “simultaneously” in the present application may be understood as at the same time point, within a period of time, or within the same cycle, and may be understood in conjunction with the context.

[0142] Elements used in the singular herein are intended to mean "one or more" rather than "one and only one" unless specifically stated otherwise.

[0143] Additionally, the terms "system" and "network" are often used interchangeably herein.

[0144] It can be understood that in each embodiment of the present application, "A and B correspond" means that B is associated with A, and B can be determined according to A. However, it should also be understood that determining B according to A does not mean determining B only according to A, and B can also be determined according to A and / or other information.

[0145] It can be understood that in the embodiments of the present application, "used to indicate" and "indicate" can include direct indication and indirect indication, and can also include explicit indication and implicit indication. When describing "a certain indication information is used to indicate A" or "indication information of A", it can include that the indication information directly indicates A or indirectly indicates A, but it does not mean that the indication information must carry A. The information indicated by a certain information is called information to be indicated. In the specific implementation process, there are many ways to indicate the information to be indicated, such as but not limited to, the information to be indicated can be directly indicated, such as the information to be indicated itself or the index of the information to be indicated. The information to be indicated can also be indirectly indicated by indicating other information, wherein there is an association relationship between the other information and the information to be indicated. It is also possible to indicate only a part of the information to be indicated, while the other parts of the information to be indicated are known or agreed in advance. For example, the indication of specific information can also be realized by means of the arrangement order of each information agreed in advance (for example, stipulated by the protocol), thereby reducing the indication overhead to a certain extent. At the same time, the common parts of each information can also be identified and uniformly indicated to reduce the indication overhead caused by indicating the same information separately. In addition, the specific indication method can also be various existing indication methods, such as but not limited to the above-mentioned indication methods and various combinations thereof. The specific details of various indication methods can refer to the prior art, and will not be repeated herein. As described above, it can be seen that, for example, when it is necessary to indicate multiple information of the same type, different indication methods may appear. In the specific implementation process, the required indication method can be selected according to specific needs. The embodiment of the present application does not limit the selected indication method. In this way, the indication method involved in the embodiment of the present application should be understood as covering various methods that can enable the party to be indicated to know the information to be indicated. The information to be indicated can be sent as a whole, or it can be divided into multiple sub-information and sent separately, and the sending period and / or sending time of these sub-information can be the same or different. The specific sending method is not limited in this application. Among them, the sending period and / or sending time of these sub-information can be pre-defined, for example, pre-defined according to the protocol, or it can be configured by the transmitting end device by sending information to the receiving end device. Among them, the information can be, for example, but not limited to, one or a combination of at least two of radio resource control signaling, MAC layer signaling and physical layer signaling.

[0146] The technical solutions of the embodiments of the present application can be applied to various communication systems, such as: universal mobile telecommunications system (UMTS) or third generation (3G) system, long term evolution (LTE) system or fourth generation (4G) system, worldwide interoperability for microwave access (WiMAX) communication system, fifth generation (5G) system, such as new generation radio access technology (NR), network integrating multiple systems, Internet of Things system, Internet of Vehicles system, and future communication systems, such as 6G system, etc. The technical solutions of the embodiments of the present application can also be applied to open access network (open RAN, O-RAN or ORAN), cloud radio access network (cloud radio access network, CRAN), or a communication network including two or more of the above networks.

[0147] For example, see Figure 1 , Figure 1 is a schematic diagram of a scenario architecture to which a communication method provided in an embodiment of the present application can be applied, such as Figure 1 As shown, the architecture may include user equipment (UE), wireless access network equipment (such as base station xNB), and core network (CN) equipment.

[0148] The functions of the core network are mainly to provide user connections, user management and service bearing. The core network can be used as a bearer network to provide an interface to the external network. Taking the 5G core network as an example, the core network may include network elements such as user plane function (UPF), access and mobility management function (AMF), session management function (SMF), etc. It may also include other network elements for slicing, authentication, billing or policy management ( Figure 1(not shown in the figure). In an embodiment of the present application, the control node of the large model can be deployed on a certain network element already existing in the core network, or can be deployed on a network element established separately in the core network, so that the network element can control the deployment and operation of the large model, etc. Specifically, the network element deployed with the control node of the large model can configure affinity operators to other communication devices deployed with the task execution node of the large model, so that it can compress the service data to be transmitted between different communication devices through the affinity operator, thereby improving the performance of the large model. Furthermore, the network element can also configure a sparse operator for the communication device deployed with another task execution node of the large model, so that it can decompress the data compressed by the affinity operator through the sparse operator to reduce information loss. It should be noted that the control node of the large model can also be deployed on a wireless access network device or a terminal device, as long as the capabilities of these devices allow the deployment of the control node, and even the control node of the large model can also be deployed on a certain communication device together with a certain task execution node. For the convenience of understanding, the present application only takes the deployment of the control node on the core network network element as an example to expand the description, and does not constitute a limitation of the present application.

[0149] In the above architecture, the core network can be a 5G core network, a 4G core network, a 6G core network, or a core network in other future networks, without specific limitation. Each network element can also be called a functional entity, which can be a network element implemented on dedicated hardware, a software instance running on dedicated hardware, or an instance of a virtualized function on an appropriate platform, for example, the above virtualization platform can be a cloud platform. Different network elements or devices can communicate through interfaces.

[0150] For the architecture between the wireless access network device and the terminal device in the embodiment of the present application, please refer to Figure 2a The hint, Figure 2a 1 is a schematic diagram of the architecture of a communication system provided in an embodiment of the present application. The communication system includes a wireless access network 100. The wireless access network 100 may include at least one wireless access network device (such as Figure 2a 110a and 110b), the wireless access network 100 may also include at least one terminal device (such as Figure 2a 120a-120j in the figure). The terminal device can be connected to the wireless access network device in a wireless manner; the terminal devices and the terminal devices, and the wireless access network devices and the wireless access network devices can be connected to each other in a wired or wireless manner. It can be understood that Figure 2a The communication system architecture is only a schematic diagram. The communication system may also include other network devices, such as wireless relay devices and wireless backhaul devices. Figure 2a Not shown in the figure.

[0151] The wireless access network equipment can be called network equipment. It is the access equipment for the terminal to access the communication system wirelessly. It is mainly responsible for all functions related to the air interface: such as wireless link maintenance function, maintaining the wireless link with the terminal, and protocol conversion of wireless link data and IP data quality monitoring; wireless resource management function, including the establishment and release of wireless links, scheduling and allocation of wireless resources, etc.; some mobility management functions, including configuring the terminal to measure, evaluating the quality of the terminal wireless link, and deciding the terminal switching between cells. In the embodiment of the present application, multiple task execution nodes (or sub-neural network models) in the large model can be deployed in different wireless access network devices (such as Figure 2a 110a and 110b in the figure), so that these access network devices can have the function of the task execution node of the large model, so that the business data of the large model can be processed. Furthermore, these access network devices can obtain the indication information of the affinity operator from the core network network element side where the large model control node is deployed, and then first compress the business data to be transmitted between different task execution nodes through the affinity operator, and then transmit the compression result to improve the performance of the large model. It should be noted that the task execution node of the large model can also be deployed on various network elements or terminal devices of the core network, as long as the capabilities of these devices allow the deployment of task execution nodes, and even the control node of the large model can also be deployed together with a task execution node on a certain communication device. For the convenience of understanding, the present application only uses the example of deploying multiple task execution nodes on different access network devices respectively to expand the explanation, and does not constitute a limitation of the present application.

[0152] The wireless access network equipment can be a base station (BS), an evolved NodeB (eNodeB, eNB for short), a transmission reception point (TRP), a next generation base station (gNB) in the fifth generation (5G) mobile communication system, a next generation base station in the sixth generation (6G) mobile communication system, a base station in a future mobile communication system, or an access node in a wireless fidelity (WiFi) system, etc.; it can also be a module or unit that completes part of the functions of a base station, for example, it can be a centralized unit (CU), a distributed unit (DU), or a radio unit (RU). The CU here completes the functions of the radio resource control protocol and the packet data convergence layer protocol (PDCP) of the base station, and can also complete the function of the service data adaptation protocol (SDAP); the DU completes the functions of the radio link control layer and the medium access control (MAC) layer of the base station, and can also complete the functions of part of the physical layer or all of the physical layer. For the specific description of the above-mentioned various protocol layers, please refer to the relevant technical specifications of the 3rd generation partnership project (3GPP). CU and DU can be set separately, or they can also be included in the same network element, such as the baseband unit (BBU). RU can be included in a radio frequency device or radio frequency unit, such as a remote radio unit (RRU), an active antenna unit (AAU) or a remote radio head (RRH). In different systems, CU, DU or RU may also have different names, but those skilled in the art can understand their meanings. For example, in the ORAN system, CU may also be referred to as O-CU (open CU), DU may also be referred to as O-DU, and RU may also be referred to as O-RU. Any of the CU (or CU-CP, CU-UP), DU, and RU in the present application may be implemented by a software module, a hardware module, or a combination of a software module and a hardware module. The wireless access network device may be a macro base station (such as Figure 2a110a), or a micro base station or an indoor station (such as Figure 2a 110b), may also be a relay node or a donor node, etc. The embodiments of the present application do not limit the specific technology and specific device form adopted by the wireless access network device.

[0153] In the embodiment of the present application, the wireless access network device may be a device for realizing the function of the wireless access network device; or it may be a device capable of supporting the wireless access network device to realize the function, such as a chip system, or a communication module, or a modem, etc., which may be installed in the wireless access network device. In the technical solution provided in the embodiment of the present application, the device for realizing the function of the wireless access network device is the wireless access network device, and the wireless access network device is a base station as an example to describe the technical solution provided in the embodiment of the present application. The base station may support networks with the same or different access technologies. The embodiment of the present application does not limit the specific technology and specific device form adopted by the wireless access network device.

[0154] The terminal device is a device with wireless transceiver function, which can send signals to the base station or receive signals from the base station. The terminal device can also be called a terminal, user equipment (UE), mobile station, mobile terminal, etc. The terminal device can be widely used in various scenarios, for example, device-to-device (D2D), vehicle to everything (V2X) communication, machine-type communication (MTC), Internet of Things (IOT), virtual reality, augmented reality, industrial control, automatic driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, smart city, etc. The terminal can be a mobile phone, a tablet computer, a computer with wireless transceiver function, a wearable device, a vehicle, an airplane, a ship, a robot, a mechanical arm, a smart home device, a train detector, a gas station sensor, etc. The embodiments of the present application do not limit the specific technology and specific device form adopted by the terminal. It can be understood that the physical device and the mobile user can be completely independent, and all information related to the mobile user can be stored in a smart card (such as a SIM card), which can be used on the physical device. In an embodiment of the present application, the terminal device side can be used to initiate a business request (such as an inference request) for a large model, and then multiple task execution nodes of the large model process the business request, and finally return the business result to the terminal device. The compression of affinity operators is involved in the business data transmission process to ensure the performance of the large model; further, the decompression of sparse operators may also be involved in the business data transmission process to reduce information loss. It should be noted that if the terminal device has sufficient capabilities, the terminal device may also be deployed with task execution nodes and / or control nodes of the large model. For ease of understanding, this application only uses the terminal device as the initiator of the large model business request, and does not deploy large model task nodes and control nodes as an example for explanation, which does not constitute a limitation of this application.

[0155] In the embodiment of the present application, the device for realizing the function of the terminal may be a terminal; or it may be a device capable of supporting the terminal to realize the function, such as a chip system, or a communication module, or a modem, etc., which may be installed in the terminal. In the embodiment of the present application, the chip system may be composed of chips, or may include chips and other discrete devices. In the technical solution provided in the embodiment of the present application, the device for realizing the function of the terminal is a terminal, and the terminal is a UE as an example to describe the technical solution provided in the embodiment of the present application. The embodiment of the present application does not limit the specific technology and specific device form adopted by the terminal device.

[0156] exist Figure 2aIn the example communication system, the base station and the terminal can be fixed or movable. The base station and the terminal can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on the water surface; they can also be deployed on airplanes, balloons, and artificial satellites. The embodiments of the present application do not limit the application scenarios of the base station and the terminal. In addition, the roles of the base station and the terminal can be relative, for example, Figure 2a The helicopter or drone 120i in the figure can be configured as a mobile base station. For the terminal 120j that accesses the wireless access network 100 through 120i, the terminal 120i is a base station; but for the base station 110a, 120i is a terminal, that is, 110a and 120i communicate through the wireless air interface protocol. Of course, 110a and 120i can also communicate through the interface protocol between base stations. In this case, relative to 110a, 120i is also a base station. Therefore, base stations and terminals can be collectively referred to as communication devices. Figure 2a 110a and 110b in the figure may be referred to as communication devices having base station functions. Figure 2a 120a-120j in the figure can be called communication devices with terminal functions. Similarly, network elements of the core network can also be called communication devices, that is, communication devices with certain specific functions of the core network.

[0157] Base stations and terminals, base stations and base stations, and terminals and terminals can communicate through authorized spectrum, through unlicensed spectrum, or through both authorized spectrum and unlicensed spectrum. They can communicate through spectrum below 6 gigahertz (GHz), through spectrum above 6 GHz, or through both spectrum below 6 GHz and spectrum above 6 GHz at the same time.

[0158] It should be noted that in some scenarios, UE can also be used to act as a base station. For example, UE can act as a scheduling entity, which provides sidelink signals between UEs in vehicle-to-everything (V2X), device-to-device (D2D) or peer to peer (P2P). In some scenarios, UE can also be used to act as a relay node. For example: UE can act as a relay device (relay) or an integrated access and backhaul (IAB) node to provide wireless backhaul services for terminal devices.

[0159] Optionally, the method provided in this application can be applied to V2X scenarios (such as the aforementioned Figure 2aThe present invention is applicable to the scenario where 120a and 120b communicate with each other), and can also be applied to the scenario where the sidelink UE-to-network relay (Sidelink UE-to-Network Relay) and the scenario where the sidelink UE-to-UE relay (Sidelink UE-to-UE Relay) are used. Figure 2b , Figure 2b Schematic diagram of the Sidelink UE-to-Network Relay scenario provided by the embodiment of the present application. Figure 2b As shown, the Sidelink UE-to-NetworkRelay scenario includes Remote UE (remote UE) and Relay UE (relay UE), and also includes a base station; the method provided in this application can be applied to the communication between Remote UE and Relay UE. Figure 2c , Figure 2c Schematic diagram of the Sidelink UE-to-UE Relay scenario provided by the embodiment of the present application. Figure 2c As shown, the Sidelink UE-to-UE Relay scenario includes a Source UE, a Relay UE, and a target UE; the method provided in the present application can be applied to communication between a Source UE and a Relay UE and / or communication between a Relay UE and a target UE.

[0160] It should be understood that the applicable scenario architecture and communication system architecture described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. A person of ordinary skill in the art can know that with the evolution of the network architecture, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0161] In the embodiments of the present application, the term "wireless communication" may also be referred to as "communication", and the term "communication" may also be described as "data transmission", "information transmission" or "transmission".

[0162] See also Figure 3 , Figure 3 is a simplified structural diagram of a UE and a base station provided in an embodiment of the present application. For simplicity, Figure 3 Only base station 110 (corresponding to Figure 2a 110a and 110b) and UE120 (corresponding to Figure 2a In the embodiment of the present invention, the structure of the base station and the UE may be more than 120a-120j). Figure 3 Parts shown, less than Figure 3The components shown may also include only Figure 3 It should be understood that Figure 3 The base station shown can adopt either a CU-DU separation architecture or a non-CU-DU separation architecture. The simplified structure of the core network side device can refer to the structure of the base station and UE, and will not be expanded separately. Figure 3 A brief introduction to each component is given below.

[0163] The base station 110 includes an interface 111 and a processor 112. The processor 112 may optionally store a program 114. The base station 110 may optionally include a memory 113. The memory 113 may optionally store a program 115. The UE 120 includes an interface 121 and a processor 122. The processor 122 may optionally store a program 124. The UE 120 may optionally include a memory 123. The memory 123 may optionally store a program 125. These components work together to provide various functions described in this application. For example, the processor 112 and the interface 111 work together to provide a wireless connection between the base station 110 and the UE 120. The processor 122 and the interface 121 work together to implement downlink transmission and / or uplink transmission of the UE 120.

[0164] The processor (e.g., processor 112 and / or processor 122) may include one or more processors and be implemented as a combination of computing devices. The processor (e.g., processor 112 and / or processor 122) may include one or more of the following: a microprocessor, a microcontroller, a digital signal processor (DSP), a digital signal processing device (DSPD), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic device (PLD), a gating logic, a transistor logic, a discrete hardware circuit, a processing circuit or other suitable hardware, firmware and / or a combination of hardware and software to perform the various functions described in this application. The processor (e.g., processor 112 and / or processor 122) may be a general-purpose processor or a dedicated processor. For example, processor 112 and / or processor 122 may be a baseband processor or a central processing unit. The baseband processor may be used to process communication protocols and communication data. The central processing unit may be used to enable the base station 110 and / or UE 120 to execute software programs and process data in the software programs.

[0165] The interface (e.g., interface 111 and / or 121) may include a device for implementing communication with one or more computer devices (e.g., UE, BS, and / or network node). In some embodiments, the interface may include wires for coupling wired connections, or pins for coupling wireless transceivers, or chips and / or pins for wireless connections. In some embodiments, the interface may include a transmitter, a receiver, a transceiver, and / or an antenna. The interface may be configured to use any available protocol (e.g., 3GPP standard).

[0166] The program in this application can be used to represent software in a broad sense. Non-limiting examples of software are program codes, programs, subroutines, instructions, instruction sets, codes, code segments, software modules, applications, software applications, etc. The program can be run in a processor and / or a computer to enable the base station 110 and / or the UE 120 to perform various functions and / or processes described in this application.

[0167] Memory (e.g., memory 113 and / or memory 123) can store data manipulated by processors 112, 122 when executing software. Memories 113, 123 can be implemented using any storage technology. For example, the memory can be any available storage medium that can be accessed by the processor and / or computer. Non-limiting examples of storage media include: RAM, ROM, EEPROM, CD-ROM, removable media, optical disk storage, magnetic disk storage media, magnetic storage devices, flash memory, registers, state memory, remote mounted storage, local or remote memory components, or any other medium that can carry or store software, data or information and can be accessed by the processor / computer.

[0168] The memory (e.g., memory 113 and / or memory 123) and the processor (e.g., processor 112 and / or processor 122) may be separately arranged or integrated together. The memory may be used to connect to the processor so that the processor can read information from the memory and store and / or write information in the memory. The memory 113 may be integrated into the processor 112. The memory 123 may be integrated into the processor 122. The processor (e.g., processor 112 and / or processor 122) and the memory (e.g., memory 113 and / or memory 123) may be arranged in an integrated circuit (e.g., the integrated circuit may be arranged in a UE or a base station or other network node).

[0169] Optionally, the components in the above-mentioned base station 110 and UE 120 may also exist in the form of various interaction modules, for example, a radio resource control (RRC) signaling interaction module, that is, a module used by the base station 110 and UE 120 to send and receive RRC signaling; a MAC signaling interaction module, that is, a module used by the base station 110 and UE 120 to send and receive medium access control-control element (MAC-CE) signaling; a PHY signaling and data interaction module, that is, a module used by the base station 110 and UE 120 to send and receive uplink / downlink control signaling, and uplink / downlink data.

[0170] The above content explains the communication system architecture that can be applied to the embodiments of the present application. In order to better understand the technical solutions of the embodiments of the present application, several terms or nouns related to the present application are briefly introduced below to facilitate understanding by technical personnel in this field.

[0171] 1. Large Model

[0172] A large model refers to a machine learning model with a large number of parameters and a complex structure, generally a deep neural network with more than 1 billion parameters. The large model can process massive amounts of data and complete various complex tasks, such as natural language processing, computer vision, speech recognition, etc. Among them, natural language processing may include machine translation, text generation, language models, etc., computer vision may include image classification, target detection, image generation, etc., and speech recognition may include speech-to-text, speech recognition, etc. The present application can solve the problem that when multiple nodes for performing tasks in a large model (or neural network models, such as a first sub-neural network model and a second sub-neural network model) are respectively deployed on different communication devices in a mobile communication network, so that these communication devices can handle the business of the large model, the overall performance of the large model is severely limited.

[0173] 2. Large language model (LLM)

[0174] At present, the more common large model should be a large language model, such as Llama, ChatGPT, etc. The large language model can understand and generate human language by learning a large amount of text data, and usually has the characteristics of deep layers, large number of parameters, and large amount of training data. In essence, the large language model is a probability-based autoregressive language model (autoregressivemodel, AR model), and the AR model generates text by predicting the next representation (token) in the text sequence. During the training process, the model processes the input sequence token by token and predicts the probability distribution of the next token; during the generation process, the model gradually generates new tokens according to the context and the generated tokens until the entire text sequence is completed. The key to the large language model includes text semantic representation based on word embedding, converters based on attention mechanisms (such as Transformer), and self-supervised learning based on predicting the next word. For ease of understanding, this application will take the large model as a large language model as an example to explain the communication method and processing flow involved in the combination of the large model and the mobile communication network in this application. Understandably, the large model in this application can be other large-scale models in addition to the large language model.

[0175] 3. Queries-per-second (QPS)

[0176] The query rate per second indicates the amount of request data (generating inference results) processed by a unit of computing resources in a unit of time. In large model scenarios, an increase in QPS means an increase in the speed of large model inference, which can provide services to more users or reduce the cost of model inference.

[0177] 4. Encoder and Decoder in Large Language Model

[0178] Encoders and decoders are usually neural network models. The difference between encoders and decoders lies in their inputs and outputs as well as their functions. In simple terms, encoders are used to encode input information, and decoders are used to generate output information. Encoders are responsible for converting inputs into features, and decoders are responsible for converting features into targets and outputting them. Specifically, the input of the encoder is the input sequence, and the output is the hidden vector representation of each position; the input of the decoder is the output of the encoder and part of the output sequence generated previously, and the output is the word generated at the next position.

[0179] 5. Operator

[0180] An operator, also called an operator, is a mapping from a function space to a function space. Furthermore, operators can be generalized to any space, such as inner product space. Broadly speaking, any operation on any function can be considered an operator operation, such as exponentiation and square root can be considered an operator operation. For example, in a network model, the operator corresponds to the calculation logic in the layer, for example: the convolution layer is an operator, and the weight summation process in the fully connected layer is an operator. Corresponding to the operator is the object being operated, generally referred to as an operand. Common operators include differential operators, gradient operators, divergence operators, etc. In an embodiment of the present application, in order to solve the problem of performance degradation when different task execution nodes of a large model are deployed on different communication devices, the network side can configure an affinity operator for the previous communication device to compress the business data to be transmitted between different communication devices, thereby improving the performance of the large model; further, in order to solve the problem that in certain scenarios, after the business data is compressed by the affinity operator, the information distribution changes, becomes dense, and causes information loss, the network side can configure a sparse operator for the next communication device to decompress the business data compressed by the affinity operator, thereby reducing information loss.

[0181] 6. Forward Propagation and Back Propagation

[0182] Forward propagation refers to the process of transferring information from the input layer to the output layer in a neural network. Specifically, in the process of forward propagation, the output of the previous layer is used as the input of the next layer, and the output of the next layer is calculated, and the calculation continues until the output layer. In this application, forward propagation mainly refers to the process of processing business data by the task execution node, such as the processing of business data by the first sub-neural network model in the second communication device and the processing of business data by the second sub-neural network model in the third communication device.

[0183] Back propagation is a method for calculating the gradient of each layer of parameters in a neural network. Usually, the partial derivatives of the loss function with respect to the weights and biases of each neuron are obtained layer by layer using the chain rule, and the gradient of the loss function with respect to the weights and bias vectors is formed, which is used as the basis for modifying the weights. In this application, back propagation is mainly used for gradient feedback during the training of affinity operators and / or sparse operators, such as the third communication device feeding back the result of the gradient calculation to the second communication device.

[0184] First, the technical problems that this application specifically aims to solve are analyzed and proposed. At present, the big model is basically deployed on the server in the data center, and there is still a lack of research on its deployment in the mobile communication network, that is, there is currently a lack of solutions for combining the big model with the mobile communication network. The applicant has made some attempts to combine the big model with the mobile communication network. Considering the large size of the big model, if the big model is considered to be deployed in the mobile communication network, the combination of the big model and the mobile communication network can be achieved by splitting the big model into multiple task execution nodes, and then deploying these task execution nodes on different communication devices in the mobile communication network, so that these communication devices have the ability to process the big model business. Figure 4 As shown, taking the large model as an example, a large language model based on a transformer, the model may include two types of transformer blocks, one is an encoder, and the other is a decoder. The number of each type of transformer block is 6 (such as encoder1-6, decoder1-6). Exemplarily, different transformer blocks are deployed on different base stations, for example, encoder1-3 is deployed on base station 1, encoder4-6 is deployed on base station 2, and decoder1-6 can be deployed on base station 3 and base station 4, respectively. After the large model service data from the UE is input into the previous block (base station 1, such as corresponding to encoder1-3) for processing, since the block outputs the data of the fully connected layer and the data dimension is large, the amount of data that needs to be transmitted to the next block (base station 2, such as corresponding to encoder4-6) will be expanded many times, for example Figure 4 As shown in the figure, the data size of the problem input from the UE is only 0.4KB, and when it is transmitted to the middle layer (between base station 1 and base station 2), the data size becomes 772KB, which is nearly 2000 times larger. In actual business scenarios, the size of business data from the UE is often larger, so the amount of data that needs to be transmitted after the middle layer is expanded is even larger (even up to the order of 10e6). However, due to the serious lack of x2 interface bandwidth between base stations, the final inference capacity (such as QPS) of the large model is limited to the order of 1e6, which seriously reduces the performance of the large model.

[0185] To this end, the present application proposes a communication method, device and readable storage medium. When multiple task execution nodes (TE, corresponding to one or more blocks) of a large model are respectively deployed on different communication devices of a mobile communication network, the network side can configure a network affinity operator (i.e., the first operator) for compressing the business data to be transmitted between multiple task execution nodes, thereby reducing the data volume of the business data that needs to be transmitted between different task execution nodes. Even when the interface bandwidth between different communication devices is limited, data transmission can be completed effectively and timely, thereby improving the performance of the large model.

[0186] For ease of understanding, the technical solution provided by this application will be described below in conjunction with more drawings.

[0187] In this application, unless otherwise specified, the same or similar parts between various embodiments or implementations can refer to each other. In each embodiment of this application, and each implementation method / implementation method / implementation method in each embodiment, if there is no special explanation and logical conflict, the terms and / or descriptions between different embodiments and each implementation method / implementation method / implementation method in each embodiment are consistent and can be referenced to each other. The technical features in different embodiments and each implementation method / implementation method / implementation method in each embodiment can be combined to form new embodiments, implementation methods, implementation methods, or implementation methods according to their inherent logical relationships. The implementation methods of this application described below do not constitute a limitation on the scope of protection of this application.

[0188] See also Figure 5a , Figure 5a is a flow chart of a communication method provided in an embodiment of the present application, which can be applied to the above Figure 1 The UE, base station and core network equipment described in the mobile communication network architecture, that is, the above Figure 1 The UE, base station and core network equipment described in the above may be used to support and execute Figure 5a Relevant steps of the method flow shown in . Optionally, the first communication device may be a core network device, an access network device or a UE, or may be a module / device (such as a chip) in a core network device, an access network device or a UE, or may be a component (such as a DU) that performs part of the functions of the core network device, the access network device or the UE. Similarly, the second communication device and the third communication device may also be a core network device, an access network device or a UE, or a module / device therein, or a component that performs part of their functions.

[0189] like Figure 5a As shown, taking the first communication device as a core network element, the second communication device as a base station 1, and the third communication device as a base station 2 as an example, the communication method includes but is not limited to the following steps:

[0190] S501: A first communication device sends first information to a second communication device, where the first information is used to indicate a first operator. Correspondingly, the second communication device receives the first information from the first communication device.

[0191] Specifically, a control node of the large model is deployed on the first communication device, and the control node can also be called a task control function (TCF) node. The large model can also include multiple task execution (TE) nodes, such as TE1 and TE2. Each task execution node can correspond to a sub-neural network model, such as TE1 can correspond to the first sub-neural network model, and TE2 can correspond to the second sub-neural network model. In an embodiment of the present application, one or more large model blocks (such as encoders and decoders) deployed on a communication device can be considered as a complete sub-neural network model, that is, a task execution node of the large model can include one or more large model blocks. Multiple sub-neural network models in the large model can be deployed on different communication devices respectively, such as TE1 (corresponding to the first sub-neural network model) is deployed on the second communication device, and TE2 (corresponding to the second sub-neural network model) is deployed on the third communication device. The first communication device can send configuration information to the second communication device to indicate the first operator, and the first operator can be used to compress the business data to be transmitted between the above-mentioned multiple sub-neural network models, such as the first operator is used to compress the business data transmitted by the second communication device to other communication devices (such as the third communication device). In other words, the control node (TCF) of the large model can send configuration information to TE1 to indicate the first operator, and the first operator can be used to compress the business data to be transmitted between the above-mentioned TE1 and TE2, thereby reducing the amount of data transmitted between different nodes. The magnitude of the business data improves the performance of the large model. In an embodiment of the present application, the operator used to compress the business data to be transmitted between multiple sub-neural network models can also be called an affinity operator, and the operator used to decompress the business data compressed by the affinity operator can be called a sparse operator.

[0192] Optionally, after the first communication device determines the deployment scheme of the large model in the mobile communication network, it can send instruction information to each communication device that needs to deploy the sub-neural network model of the large model, instructing these communication devices to load the corresponding sub-neural network models respectively, so that these communication devices have the ability to process the large model business. Exemplarily, after the first communication device determines that TE1 (i.e., corresponding to the first sub-neural network model) needs to be deployed on the second communication device, and TE2 (i.e., corresponding to the second sub-neural network model) needs to be deployed on the third communication device, it can send the seventh information to the second communication device and send the eighth information to the third communication device. Among them, the seventh information can include the first parameter information, and the seventh information can be used to instruct the second communication device to load the first sub-neural network model based on the first parameter information; the eighth information can include the second parameter information, and the eighth information can be used to instruct the third communication device to load the second sub-neural network model based on the second parameter information. Optionally, the first parameter information and the second parameter information can respectively include the model structure of the corresponding sub-neural network model (such as how many layers the model has and the number of neurons in each layer), the loading platform for loading the sub-neural network model, and the management resource requirements (how much memory and storage are required).

[0193] S502: The second communication device processes the acquired first service data, and compresses the processing result based on the first operator.

[0194] Specifically, the second communication device is deployed with a large model TE1 (i.e., corresponding to the first sub-neural network model). After obtaining the first business data, the second communication device can first input the first business data into the first sub-neural network model for processing and output the processing result (i.e., the second business data), and then compress the second business data through the first operator to obtain the third business data. Among them, the data volume of the third business data is smaller than the data volume of the second business data. Optionally, the first business data obtained by the second communication device may be an inference request initiated by the UE, or may be initiated by the second communication device itself, or may be an inference request initiated by other communication devices, which is not specifically limited here. Optionally, the second communication device may determine the first operator before obtaining the first business data, or may determine the first operator after obtaining the first business data, which is not specifically limited here.

[0195] S503: The third communication device receives and processes the compression result sent by the second communication device.

[0196] Specifically, after completing the compression of the business data, the second communication device can send the compression result to the third communication device, and the third communication device can further process the compression result based on the second sub-neural network model. Optionally, if the third communication device is deployed with the last task execution node in the large model (such as decoder6 in the large language model), then after the third communication device completes the processing of the compression result, it can return the processing result to the communication device that initiated the reasoning request as a response to the reasoning request. Exemplarily, if the UE is the initiator of the reasoning request, the third communication device can return the processing result to the UE as a response to the reasoning request. Furthermore, if the UE is not within the service range of the third communication device (base station 2), the third communication device can first find the base station currently providing services to the UE (which can be base station 1, i.e., the second communication device, or other base stations), and then it can forward the processing result to the UE. If the third communication device is not deployed with the last task execution node of the large model, but an intermediate node (such as decoder3 in the large language model), then the third communication device needs to continue to transmit the processing results to the subsequent communication device (such as the communication device deployed with decoder4 in the large language model), and the subsequent communication device will process the corresponding functions until the service data processing is completed and transmitted to the communication device where the last task execution node of the large model is deployed. Finally, after the communication device completes the processing of the service data, it returns the response of the inference request to the UE. Optionally, assuming that the starting node of the large model is deployed on the second communication device, the UE that initiates the inference request may be within the service range of the second communication device or may not be within the service range of the second communication device. If the UE is within the service range, the UE can directly send the inference request to the second communication device; if the UE is not within the service range, the UE can send the inference request to the access network device that provides services for it, and then the access network device forwards the inference request to the second communication device for processing.

[0197] Optionally, the principle of the affinity operator compressing the service data in step S502 may refer to the following methods:

[0198] (1) Affinity operators can reduce the repetition and redundancy in business data information through a special encoding method to achieve data compression. For example, if the content of a piece of data is 000111…111000 (with 20 1s in the middle), the data will be long if all of it is written out. Through encoding, the 20 1s in the middle can be represented by less data, such as 000 (20 1s) 000, thus completing the compression.

[0199] (2) The affinity operator can be a low-dimensional network affinity layer. The lower the dimension of the affinity layer, the stronger the compression ability of the affinity operator. The business data output after block (corresponding sub-neural network model) processing includes information of multiple neurons. After the output business data is processed by the affinity operator, only a part of the multiple neurons have values ​​greater than the preset threshold value, while the values ​​of other neurons are less than or equal to the preset threshold value. For example, only a part of the multiple neurons have values, and the values ​​of other neurons are 0. At this time, the second communication device only needs to transmit this part of the valuable neuron information to the third communication device, and the other 0 value neuron information is not transmitted, thereby completing the compression. For example, if Figure 5b As shown in the structure, the second business data output after the first business data is processed by block1 includes the information of M (for example, M is 10) neurons. The second business data is compressed by the affinity operator to obtain the third business data and sent to block2. The third business data includes the information of N (for example, N is 2) neurons. These N neurons are part of the M neurons, and N is a positive integer less than M. Assuming that M neurons are the number of all neurons, for the block direct connection method, Figure 5b In the example, the business data transmitted from block 1 to block 2 needs to include all the information of these M neurons. However, for the method of adding affinity operators between blocks, the transmitted business data only needs to include the information of N neurons among these M neurons, that is, part of the neuron information among all neurons. For ease of understanding, we can assume that the affinity operator matrix is ​​W e , the output of block1 is X∈R M , then the output after affinity operator processing is The neuron values ​​of the first N dimensions of Y are valid and passed backward, where X represents the neuron output of block 1, the dimension of X is M, and R M represents an M-dimensional real number, where M is the dimension of the output neuron of block 1, Y represents the neuron output of the affinity operator, and δ e represents the activation function corresponding to the affinity operator, Represents matrix multiplication.

[0200] (3) Combination of method (1) and method (2). For example, the affinity operator can first reduce the number of neurons in the business data, and then reduce the duplication and redundancy in the information of each neuron to achieve a compression effect. The above two methods can be combined with each other, which will not be further described here.

[0201] In a possible implementation, the affinity operator can be determined based on the model type and model splitting point of the large model. Exemplarily, the network side can first preset an affinity operator matching table, as shown in Table 1, wherein a large model can be configured with one or more splitting points, and the number of configurable splitting points for different types of large models can be the same or different; for the same large model, when selecting different splitting points for model splitting, the corresponding affinity operators can be configured respectively.

[0202] Table 1

[0203] Model Type Split Point Affinity Operator Chatgpt 1 Affinity Operator #1 Chatgpt 2 Affinity Operator #2 Chatgpt 3 Affinity Operator #3 Pangu 1 Affinity Operator #4 Pangu 2 Affinity Operator #5

[0204] For example, Figure 5c As shown, a large model A may include 6 Transformer blocks (all decoders), of which there are 3 split points. When the large model A is to be deployed in the mobile communication network, if the first communication device (control node) selects split point 1 as the location for splitting and deploying the large model, the task execution node (TE1) deployed on the second communication device includes decoders 1-2 (corresponding to the first sub-neural network model), and the task execution node (TE2) deployed on the third communication device includes decoders 3-6 (corresponding to the second sub-neural network model). The second communication device can first use affinity operator #1 to compress the service data, and then send the compression result to the third communication device for processing. Similarly, if the first communication device (control node) chooses to split the large model at split point 2, decoders 1-3 are deployed on the second communication device, and decoders 4-6 are deployed on the third communication device. The second communication device can first use affinity operator #2 to compress the service data, and then send the compression result to the third communication device for processing. For ease of understanding, in the embodiments of the present application, the deployment scheme in which the large model is split into 2 nodes is used as an example for explanation. It is understandable that the large model can also be split into more nodes, such as 3, 4 or even more nodes, which can be deployed on different communication devices in the mobile communication network. Optionally, the number of split points that can be configured for a large model can be determined according to the specific scale of the large model. For example, a large model with a parameter scale of more than 1 trillion can be configured with 4 split points, and a large model of a smaller scale can be configured with 2 or 3 split points. The control node can select 1 split point from multiple split points to split the large model into two parts, or select 2 split points to split the large model into 3 parts, that is, select P split points to split the large model into P+1 parts.

[0205] As for the splitting point at which the large model should be split, it can be determined according to the capabilities of each communication device that needs to deploy the large model in the mobile communication network. The example of the large model A including three splitting points is still used for explanation. The specific splitting point to be selected to split and deploy the large model can be determined according to the capabilities of the second communication device and the third communication device. If the capability of the second communication device is less than that of the third communication device, the second communication device cannot carry 3 or more decoders, and the third communication device can carry 4 or more decoders, then the large model can be split at splitting point 1, so that decoders 1-2 are deployed on the second communication device, and decoders 3-6 are deployed on the third communication device; if the capabilities of the second communication device and the third communication device are equivalent and both can carry 3 decoders, then the large model can be split at splitting point 2; or, if the capability of the second communication device is greater than that of the third communication device, the second communication device can carry 4 decoders, and the third communication device can carry 2 decoders, then the large model can be split at splitting point 3.

[0206] Further, in step S501, the first communication device may directly or indirectly instruct the second communication device to use the first operator to compress the service data through the first information. Exemplarily, for the direct indication method, the first communication device may pre-store an affinity operator matching table locally (refer to Table 1), and when the first communication device determines the deployment scheme of the large model in the mobile communication network, the type of large model to be deployed (i.e., which large model) and the model splitting point may be known, so that the affinity operator matching table may be queried based on this information to determine which first operator (affinity operator) the second communication device should use, and then directly indicate the first operator through the first information, i.e., the first information may include the first operator. For the indirect indication method, the second communication device may pre-store an affinity operator matching table locally, and after determining the type of large model to be deployed and the model splitting point, the first communication device informs the second communication device of the large model type and the model splitting point through the first information, and then the second communication device determines the corresponding first operator by looking up the table based on the large model type and the model splitting point. Furthermore, the above-mentioned affinity operator #1 can also be a set, which can include multiple affinity operators, that is, a large model type and a large model split point can correspond to multiple affinity operators, and the first communication device or the second communication device can determine an affinity operator (as the first operator) for business data compression from these multiple affinity operators according to the current data transmission rate between the second communication device and the third communication device.

[0207] In one possible implementation, the network side can reconfigure an affinity operator for the second communication device, which can be applied to the case where the data transmission rate between the second communication device and the third communication device changes, and the change value is greater than or equal to a certain threshold. In other words, the control node can reconfigure an affinity operator for two task execution nodes, and the transmission bandwidth change value between the two task execution nodes may exceed a certain threshold. In this way, the new transmission bandwidth can be adapted to maximize the performance of the large model. Optionally, the first communication device can send a fifth message to the second communication device to indicate the third operator, and the compression rate of the third operator is different from that of the first operator. Exemplarily, still taking the above Table 1 as an example, the above affinity operator #1 can be a set, which can include multiple affinity operators. After determining the model type and split point, the network side can determine that a certain affinity operator in the affinity operator #1 set needs to be used to compress the service data between the second communication device and the third communication device. Furthermore, the network side can determine the corresponding affinity operator (i.e., the first operator) from the affinity operator #1 set according to the current data transmission rate between the second communication device and the third communication device, and then instruct the second communication device to use the affinity operator to compress the service data. Afterwards, when the data transmission rate change value between the second communication device and the third communication device exceeds the preset threshold, the network side can reselect an affinity operator (i.e., the third operator) from the affinity operator #1 set, and instruct the second communication device to use the reselected affinity operator to compress the service data. For example, when the data transmission rate between the second communication device and the third communication device increases and exceeds a certain threshold, the network side can select an affinity operator with a smaller compression capability from the affinity operator #1 set to compress the service data and reduce information loss; when the data transmission rate decreases and exceeds a certain threshold, the network side can select an affinity operator with a larger compression capability from the affinity operator #1 set to compress the service data and minimize the performance degradation of the large model.

[0208] The above shows a method of determining an affinity operator by querying a preconfigured affinity operator matching table based on a large model type and a split point. In another method, a suitable affinity operator can also be trained based on a training algorithm, so that the second communication device can compress the service data through the trained affinity operator. This method can be applied to those large models that are not defined in advance or split methods that are not defined in advance, and the affinity operator cannot be determined by table lookup.

[0209] For example, see Figure 6a , Figure 6a : is a flow chart of an affinity operator training method provided in an embodiment of the present application. The method may include but is not limited to the following steps:

[0210] S601: The first communication device sends third information. Correspondingly, the second communication device receives the third information from the first communication device.

[0211] Specifically, the first communication device can configure an initial operator for the second communication device through the third information, and instruct the second communication device to start training for the initial operator. This initial operator can be used to compress the service data transmitted by the second communication device to other communication devices (such as the third communication device).

[0212] S602: The first communication device sends fourth information. Correspondingly, the third communication device receives the fourth information from the first communication device.

[0213] Specifically, the first communication device can configure a loss function for the third communication device through the fourth information, and instruct the third communication device to train the initial operator based on the loss function. The first operator in the above text can be obtained based on the training of the loss function and the initial operator.

[0214] The embodiment of the present application does not limit the order between S601 and S602.

[0215] S603: The second communication device processes the acquired first service data and outputs a first result and a second result. Correspondingly, the third communication device receives the first result and the second result output by the second communication device. Optionally, the first service data may be an inference request initiated by the UE.

[0216] Exemplarily, after the first communication device configures the initial operator for the second communication device, the second communication device may not directly enable the initial operator to compress the service data, but first enable the training function, and then enable the trained operator (which may be the first operator mentioned above) after the training result converges. Exemplarily, in the training stage, the UE's reasoning request can obtain two outputs after being input into the second communication device for processing, such as Figure 6b As shown, one of the outputs is not processed by the initial operator, but is directly output after being processed by block1 (corresponding to the first sub-neural network model) (corresponding to the first result). The data obtained on the subsequent path of this output is used as the label for training; the other output is first processed by block1 and then processed by the initial operator, corresponding to the second result. Then, the second communication device outputs these two processing results to the third communication device. Among them, the process of inputting the inference request to the first sub-neural network model for processing is also the process of forward propagation.

[0217] S604: The third communication device obtains a third result and a fourth result based on the first result and the second result.

[0218] Specifically, the third communication device inputs the first result into the second sub-neural network model for processing (which is also the process of forward propagation) to obtain the third result, and inputs the second result into the second sub-neural network model for processing (which is also the process of forward propagation) to obtain the fourth result. Optionally, since the third result has not been processed by the initial operator, the third communication device (which corresponds to the last task execution node of the large model at this time) can feed back the third result as a processing result to the UE as a response to the UE's reasoning request. Optionally, in order to ensure the smooth progress of the training process, after the training is started, the output of the second communication device after processing the business data (i.e., the first result and the second result) can be accompanied by an identifier, so that the third communication device can determine that these data are used to train the initial operator according to the identifier. Similarly, the gradient calculation results fed back by the subsequent third communication device can also be accompanied by an identifier, so that the second communication device can determine through the identifier that the initial operator is tuned according to these data.

[0219] S605: The third communication device performs gradient calculation on the third result and the fourth result based on the loss function, and feeds back the calculation result to the second communication device.

[0220] Specifically, after the third communication device feeds back the calculation result of the gradient, the second communication device can tune (update) the initial operator according to the calculation result, and then repeat the above training process (corresponding to steps S603-S605) using the tuned initial operator until the training result converges and can meet the use requirements. Optionally, when repeating the above training process, the input of the second communication device can be the original inference request initiated by the UE, or it can be a new inference request subsequently initiated by the UE, which is not specifically limited here. Exemplarily, when the training results converge, the second communication device can first send a convergence notification to the first communication device, and the notification can carry information about the trained initial operator (which can be the first operator in the previous text); then, the first communication device can send an indication message to instruct the second communication device to end the training, and instruct the second communication device to subsequently use the trained initial operator to compress the service data (see Figure 5a The process of steps S501-S503 in the above example improves the performance of large models.

[0221] In order to solve the problem of performance degradation of a large model caused by multiple nodes of the large model being deployed on different communication devices, the present application proposes a scheme to compress the business data to be transmitted between different task execution nodes through an affinity operator configured on the network side to improve the performance of the large model. Furthermore, the present application proposes that while the network side configures an affinity operator for a task execution node of the large model (such as deployed on the second communication device), it can also configure a sparse operator for another task execution node of the large model (such as deployed on the third communication device), that is, the network side configures a pair of affinity operators and sparse operators, which are used by the second communication device and the third communication device respectively, so that the third communication device can decompress the business data compressed by the affinity operator through the sparse operator, thereby reducing information loss while improving the performance of the large model.

[0222] like Figure 7a As shown, the first communication device can send the first information to the second communication device to instruct the second communication device to use the first operator to compress the service data (such as step S701), and can also send the second information to the third communication device to instruct the third communication device to use the second operator to decompress the service data compressed by the first operator (such as step S702); subsequently, after the second communication device obtains the inference request (i.e., the first service data) from the UE, it can first input it into the first sub-neural network model (corresponding to TE1) for processing (called forward propagation) to obtain the second service data, and then use the first operator to compress the second service data, and send the compressed service data (i.e., the third service data) to the third communication device (such as step S703). After receiving the compressed third service data, the third communication device can first decompress the third service data through the second operator to obtain the fourth service data, and then input the decompressed fourth service data into the second sub-neural network model (corresponding to TE2) for processing (called forward propagation) to obtain the fifth service data, and finally return the fifth service data to the UE as a response to the inference request.

[0223] Optionally, the decompression of the business data by the sparse operator may be the inverse process of the compression of the business data by the affinity operator.

[0224] For example, the principle of decompressing business data by the sparse operator can refer to the following methods:

[0225] (1) As the affinity operator uses special encoding to reduce the duplication and redundancy of business data information, the sparse operator can restore the compressed business data through a coding method that is opposite to the affinity operator. For example, a piece of data compressed by the affinity operator with the content of 000 (20 1s) 000 can be decompressed and restored to 000111…111000 (with 20 1s in the middle).

[0226] (2) Regarding the way that affinity operators achieve compression by reducing the number of neurons in business data, after the business data is processed by affinity operators, the information distribution of neurons changes and becomes more dense. In other words, after being processed by affinity operators, the mutual information between input and output may be greatly lost, affecting the accuracy of the large model. In view of this situation, sparse operators can re-adjust the information distribution of all neurons based on the information of some neurons to reduce the degree of density. For example, after compression by affinity operators, only some neurons (5) of the 10 neurons originally output by the first sub-neural network model have values, and the values ​​of the other neurons (the other 5) are 0. The second communication device only needs to transmit the neuron information with values ​​to the third communication device, while the other 0-value neuron information is not transmitted. Since the output of the first sub-neural network model and the input of the second sub-neural network model should be of the same dimension, if the output dimension of the first sub-neural network model is 10 neurons, the corresponding input dimension of the second sub-neural network model should also be 10 neurons. After being compressed by the affinity operator, only the neurons in the first five dimensions of the 10 neurons may have values ​​and be transmitted to the third communication device. Therefore, when input into the second sub-neural network model, the neurons in the first five dimensions may have values, and the values ​​of the neurons in the last five dimensions may be 0, and the information density is concentrated on the neurons in the first five dimensions. Then, before the third communication device inputs the business data (i.e., the third business data) output by the second communication device into the second sub-neural network model, it can first use the sparse operator to decompress it, readjust the information distribution of the 10 neurons, and reduce the degree of information density. Optionally, for this method of reducing the number of neurons, the goal of the affinity operator and the sparse operator is not the losslessness of the previous and next information H(X), but the losslessness of the mutual information I(X,Y) of the input and output. Therefore, the paired use of the affinity operator and the sparse operator can bring relatively large compression efficiency and help improve the performance of large models.

[0227] For example, Figure 7bAs shown in the structure, the second business data output after the first business data is processed by block1 includes the information of M (for example, M is 10) neurons. The second business data is compressed by the affinity operator to obtain the third business data and then sent to block2. The third business data includes the information of N (for example, N is 5) neurons. These N neurons are part of the M neurons, and N is a positive integer less than M. For the method in which there are only affinity operators between blocks, the transmitted business data only needs to include the information of N neurons. When inputting the next block, only these N neurons have values ​​(gray part). For the method in which a sparse operator is added, the third business data is processed by the sparse operator to obtain the fourth business data. The fourth business data includes the system information of K (for example, K is 8) neurons. Then, when the fourth business data is input into the second sub-neural network model, the neurons in 8 dimensions of the 10 neurons have values, and the values ​​of the neurons in the other 2 dimensions are 0, where K is a positive integer greater than N and less than or equal to M. For ease of understanding, it can be assumed that the sparse operator matrix is ​​W d , the output of block1 is X∈R M , then the output after affinity operator processing is The output after processing by the sparse operator is Among them, X represents the neuron output of block1, R M represents an M-dimensional real number, where M is the dimension of the output neuron of block 1, Y represents the neuron output of the affinity operator, and the neuron values ​​of the first N dimensions of Y are valid. N represents the value of the neuron in the first N dimensions, δ e represents the activation function corresponding to the affinity operator, δ d represents the activation function corresponding to the sparse operator, represents matrix multiplication, Z represents the neuron output of the sparse operator, the neuron values ​​of the first K dimensions of Z are valid, and K is a positive integer greater than N and less than or equal to M.

[0228] (3) The combination of method (1) and method (2) can refer to the description of the above method (1) and method (2) to reverse the process of affinity operator method (3), which will not be further explained here.

[0229] In a possible implementation, the above-mentioned sparse operator can be determined together with the affinity operator based on the model type of the large model and the model splitting point. Exemplarily, the network side can first preset an operator pair matching table, as shown in Table 2. Among them, a large model can be configured with one or more splitting points, and when different splitting points are selected for model splitting, corresponding operator pairs (including affinity operators and sparse operators) can be configured respectively.

[0230] Table 2

[0231] Model Type Split Point Affinity Operator & Sparse Operator Llama 1 Operator pair #1 Llama 2 Operator Pair #2

[0232] Exemplarily, the large model Llama can have two split points. If the first communication device (control node) selects split point 1 as the location for splitting and deploying the large model, then after the task execution node (TE1, corresponding to the first sub-neural network model) deployed on the second communication device processes the service data, the second communication device can first compress the processing result through the affinity operator in operator pair #1, and then transmit the compressed data to the third communication device. After receiving the service data sent by the second communication device, the third communication device can first decompress the service data through the sparse operator in operator pair #1, and then hand over the decompressed service data to the task execution node (TE2, corresponding to the second sub-neural network model) for processing.

[0233] Optionally, the first communication device can directly or indirectly instruct the third communication device to use the second operator (i.e., the sparse operator) for decompression through the second information. The specific instruction method can refer to the relevant description of the first communication device instructing the second communication device to use the first operator in the above step S501, which will not be elaborated here.

[0234] In one possible implementation, when the data transmission rate between the second communication device and the third communication device changes, and the change value is greater than or equal to a certain threshold, the network side can reconfigure an affinity operator (such as the third operator) for the second communication device, and accordingly, the network side can also reconfigure a sparse operator corresponding to the new affinity operator for the third communication device. For example, the first communication device sends the sixth information to the third communication device; the sixth information is used to indicate the fourth operator, and the fourth operator is used to decompress the service data compressed by the third operator, wherein the fourth operator is a new sparse operator and the third operator is a new affinity operator. It should be noted that the decompression rates of the new sparse operator and the old sparse operator can be different or the same, and are not specifically limited here.

[0235] Similar to the method of determining the affinity operator, in a possible implementation method, suitable affinity operators and sparse operators can also be trained based on the training algorithm, so that the second communication device can compress the service data through the trained affinity operator, and the third communication device can decompress the service data compressed by the affinity operator through the trained sparse operator. This method can be applied to those large models that are not defined in advance or the splitting method that is not defined in advance, and the affinity operator and sparse operator cannot be determined by table lookup.

[0236] Optionally, the training process of the affinity operator and the training process of the sparse operator can be performed together or separately. For example, the affinity operator (such as Figure 6a The process shown in the figure) and then train the sparse operator (see Figure 6a For ease of understanding, the following is an exemplary description of training an affinity operator together with a sparse operator:

[0237] See also Figure 8a , Figure 8a : is a flowchart of an operator pair training method provided in an embodiment of the present application. The method may include but is not limited to the following steps:

[0238] S801: The first communication device sends third information. Correspondingly, the second communication device receives the third information from the first communication device.

[0239] Specifically, the first communication device may configure an initial affinity operator for the second communication device through the third information, and instruct the second communication device to start training for the initial affinity operator.

[0240] S802: The first communication device sends fourth information. Correspondingly, the third communication device receives the fourth information from the first communication device.

[0241] Specifically, the first communication device may configure a loss function and an initial sparse operator for the third communication device through the fourth information, and instruct the third communication device to perform training on the initial affinity operator and the initial sparse operator based on the loss function.

[0242] The embodiment of the present application does not limit the order between S801 and S802.

[0243] S803: The second communication device processes the acquired first service data and outputs a first result and a second result (refer to step S603). Correspondingly, the third communication device receives the first result and the second result output by the second communication device. Optionally, the first service data may be an inference request initiated by the UE.

[0244] S804: The third communication device obtains a third result, a fourth result, a fifth result and a sixth result based on the first result and the second result.

[0245] For example, reference Figure 8bAs shown, the third communication device inputs the first result into the second sub-neural network model for processing (also the process of forward propagation) to obtain the third result (which can be used as a training label), and inputs the second result into the second sub-neural network model for processing (also the process of forward propagation) to obtain the fourth result. The first result is first processed by the initial sparse operator and then input into the second sub-neural network model for processing to obtain the fifth result. The second result is first processed by the initial sparse operator and then input into the second sub-neural network model for processing to obtain the sixth result. Optionally, since the third result has not been processed by the initial affinity operator and the initial sparse operator, the third communication device (corresponding to the last task execution node of the large model at this time) can feed back the third result as the processing result to the UE as a response to the UE's reasoning request.

[0246] S805: The third communication device performs gradient calculation on the third result, the fourth result, the fifth result and the sixth result based on the loss function, and feeds back the calculation results to the second communication device.

[0247] Specifically, after the third communication device obtains the result of the gradient calculation, the result of the gradient calculation can be fed back to the second communication device and the initial sparse operator can be tuned (updated). The second communication device can also tune (update) the initial affinity operator according to the result of the gradient calculation, and then use the tuned initial affinity operator and the tuned initial sparse operator to repeat the above training process (corresponding to steps S803-S805) until the training result converges and can meet the use requirements. Optionally, when repeating the above training process, the input of the second communication device can be the original inference request initiated by the UE, or a new inference request subsequently initiated by the UE, which is not specifically limited here. Exemplarily, when the training results converge, the second communication device and the third communication device may respectively send a convergence notification to the first communication device, and the notification may respectively carry information of the trained affinity operator and the trained sparse operator; then, the first communication device may send an indication message to instruct the second communication device and the third communication device to end the training, and instruct the second communication device to subsequently use the trained affinity operator to compress the service data, and instruct the third communication device to use the trained sparse operator to decompress the service data compressed by the affinity operator ( Figure 8a Not shown, please refer to Figure 6a description).

[0248] In summary, in the above embodiments, when multiple task execution nodes (TE, corresponding to one or more blocks) of the large model are respectively deployed on different communication devices of the mobile communication network, the network side can configure a network affinity operator (i.e., the first operator) to compress the business data to be transmitted between the multiple task execution nodes, reduce the data volume of the business data to be transmitted between different task execution nodes, and effectively and timely complete the data transmission even when the interface bandwidth between different communication devices is limited, thereby improving the performance of the large model. Furthermore, while the network side configures an affinity operator for a task execution node of the large model (such as deployed on the second communication device), it can also configure a sparse operator for another task execution node of the large model (such as deployed on the third communication device), so that the third communication device can decompress the business data compressed by the affinity operator through the sparse operator, thereby reducing information loss while improving the performance of the large model.

[0249] The above content describes the method provided by the present application. In order to facilitate the implementation of the above scheme of the embodiment of the present application, the embodiment of the present application also provides corresponding devices or equipment.

[0250] The present application divides the functional modules of the device according to the above method embodiment. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. The device of the embodiment of the present application will be described below in conjunction with the accompanying drawings.

[0251] See also Fig. 9 , Fig. 9 It is a structural diagram of a communication device provided in an embodiment of the present application. The communication device 10 can be used to implement the functions of the first communication device, the second communication device or the third communication device involved in any of the above-mentioned method embodiments, and thus can also achieve the beneficial effects possessed by the above-mentioned method embodiments.

[0252] like Fig. 9 As shown, the communication device 10 may include a sending unit 100, a receiving unit 101 and a processing unit 102. In a possible design, when the communication device 10 is used to implement the functions of the above-mentioned first communication device, the functions of each unit are as follows:

[0253] The sending unit 100 is used to send first information to a second communication device in the mobile communication network; the second communication device is deployed with a first sub-neural network model among the multiple sub-neural network models, the first information is used to indicate a first operator, the first operator is used to compress the service data transmitted by the second communication device to other communication devices, and the other communication devices are deployed with other sub-neural network models among the multiple sub-neural network models except the first sub-neural network model. Optionally, the receiving unit 101 can be used to receive various types of information sent by other communication devices; the processing unit 102 can be used to generate various types of information sent by the sending unit 100, or to process various types of information received by the receiving unit 101.

[0254] In a possible implementation, the sending unit 100 is further configured to:

[0255] Sending second information to a third communication device in the mobile communication network; the third communication device is deployed with a second sub-neural network model among the multiple sub-neural network models, the second information is used to indicate a second operator, and the second operator is used to decompress the business data compressed by the first operator.

[0256] In a possible implementation, the first operator and / or the second operator are determined based on the model type and the model splitting point of the large model, or the first operator and / or the third operator are determined based on a training algorithm.

[0257] In a possible implementation, the sending unit 100 is further configured to:

[0258] Sending third information to the second communication device; the third information includes an initial operator, and the initial operator is used to compress the service data transmitted by the second communication device to the other communication device;

[0259] Sending fourth information to a third communication device in the mobile communication network; the fourth information includes a loss function, and the first operator is obtained based on training of the loss function and the initial operator.

[0260] In a possible implementation, the sending unit 100 is further configured to:

[0261] Sending fifth information to the second communication device; the fifth information is used to indicate a third operator, the third operator is used to compress the service data transmitted by the second communication device to the other communication devices, and the compression rate of the third operator is different from the compression rate of the first operator.

[0262] In a possible implementation, the sending unit 100 is further configured to:

[0263] Sending sixth information to a third communication device in the mobile communication network; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the service data compressed by the third operator.

[0264] In a possible implementation, the sending unit 100 is further configured to:

[0265] Sending seventh information to the second communication device; the seventh information includes first parameter information of the large model, and the seventh information is used to instruct the second communication device to load the first sub-neural network model of the large model based on the first parameter information;

[0266] Sending an eighth information to a third communication device in the mobile communication network; the eighth information includes the second parameter information of the large model, and the eighth information is used to instruct the third communication device to load the second sub-neural network model of the large model based on the second parameter information.

[0267] In a possible design, when the communication device 10 is used to implement the functions of the second communication device, the functions of each unit are as follows:

[0268] A receiving unit 101 is configured to receive first information sent by a first communication device in the mobile communication network; the first information is used to indicate a first operator; the large model further includes a control node, and the control node is deployed on the first communication device;

[0269] The processing unit 102 is used to obtain first business data, and input the first business data into the first sub-neural network model for processing to obtain second business data;

[0270] The sending unit 100 is used to send third business data to a third communication device in the mobile communication network; the second sub-neural network model in the large model is deployed on the third communication device, and the third business data is obtained by compressing the second business data based on the first operator, and the data volume of the third business data is smaller than the data volume of the second business data.

[0271] In a possible implementation, the first operator is determined based on a model type and a model splitting point of the large model, or the first operator is determined by the second communication device and the third communication device based on a training algorithm.

[0272] In a possible implementation, the receiving unit 101 is further configured to:

[0273] Receive third information sent by the first communication device; the third information includes an initial operator, the third information is used to instruct the second communication device to train the initial operator, the initial operator is used to compress the service data transmitted by the second communication device to other communication devices, and the first operator is obtained based on the training of the initial operator.

[0274] In one possible implementation, the data volume of the third business data is smaller than the data volume of the second business data, including: the second business data includes information of M neurons, the third business data includes information of N neurons, the N neurons are some of the M neurons, and N is a positive integer less than M.

[0275] In a possible implementation, the receiving unit 101 is further configured to:

[0276] Receive fifth information sent by the first communication device; the fifth information is used to indicate a third operator; the third operator is used to compress business data transmitted by the second communication device to other communication devices, the compression rate of the third operator is different from the compression rate of the first operator, and the change value of the data transmission rate between the second communication device and the third communication device is greater than or equal to a preset threshold.

[0277] In a possible implementation, the receiving unit 101 is further configured to:

[0278] receiving seventh information sent by the first communication device; the seventh information includes first parameter information of the large model;

[0279] The first sub-neural network model of the large model is loaded based on the first parameter information.

[0280] In a possible implementation, the first business data is a business request sent by a terminal device, and a response to the first business data is obtained based on processing of the third business data by the second sub-neural network model.

[0281] In a possible design, when the communication device 10 is used to implement the functions of the third communication device, the functions of each unit are as follows:

[0282] A receiving unit 101 is configured to receive second information sent by a first communication device in the mobile communication network; the second information is used to indicate a second operator; the large model further includes a control node, and the control node is deployed on the first communication device;

[0283] The receiving unit 101 is further configured to receive third service data sent by a second communication device in the mobile communication network; the second communication device is configured to have a first sub-neural network model in the large model deployed thereon;

[0284] The processing unit 102 is used to input the fourth business data into the second sub-neural network model for processing to obtain fifth business data;

[0285] The sending unit 100 is used to send the fifth business data; the fourth business data is obtained by decompressing the third business data based on the second operator; the data volume of the third business data is smaller than the data volume of the fourth business data.

[0286] In a possible implementation, the second operator is determined based on a model type and a model splitting point of the large model, or the second operator is determined by the first communication device and the second communication device based on a training algorithm.

[0287] In a possible implementation, the receiving unit 101 is further configured to:

[0288] Receive fourth information sent by the first communication device; the fourth information includes a loss function, the first operator of the second communication device is obtained based on the training of the loss function, and the first operator is used to compress the service data transmitted by the second communication device to other communication devices.

[0289] In a possible implementation, the data volume of the third business data is smaller than the data volume of the fourth business data, including: the third business data includes information of N neurons, the fourth business data includes information of K neurons, the N neurons are some of the K neurons, and N is a positive integer less than K.

[0290] In a possible implementation, the receiving unit 101 is further configured to:

[0291] Receive sixth information sent by the first communication device; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the business data compressed by the third operator.

[0292] In a possible implementation, the receiving unit 101 is further configured to:

[0293] receiving eighth information sent by the first communication device; the eighth information including second parameter information of the large model;

[0294] The second sub-neural network model of the large model is loaded based on the second parameter information.

[0295] In a possible implementation, the sending unit 100 is specifically configured to:

[0296] The fifth service data is sent to the terminal device, where the fifth service data is a response to the service request initiated by the terminal device.

[0297] It should be noted that the functions of each functional unit / module in the communication device described in the embodiments of the present application can be found in the relevant description in the above method embodiments and will not be repeated here.

[0298] It can be understood that the specific descriptions of the receiving unit, sending unit and processing unit shown in the above-mentioned device embodiment are only examples. For the specific functions or execution steps of the receiving unit, sending unit and processing unit, reference can be made to the description of any of the above-mentioned method embodiments, which will not be described in detail here.

[0299] The above describes the communication device of the embodiment of the present application, and the following describes possible product forms of the communication device. It should be understood that any device having the above Fig. 9 Any form of product with the functions of the communication device described above falls within the protection scope of the embodiments of the present application. It should also be understood that the following description is only an example and does not limit the product form of the communication device of the embodiments of the present application to this.

[0300] In a possible implementation, the above Fig. 9 In the communication device shown, the processing unit 102 may be one or more processors; the sending unit 100 and the receiving unit 101 may be transceivers; or, the sending unit 100 may be a transmitter and the receiving unit 101 may be a receiver. In the embodiment of the present application, the processor and the transceiver may be coupled, etc., and the embodiment of the present application does not limit the connection method between the processor and the transceiver. In the process of executing the above method, the process of sending information in the above method can be understood as the process of outputting the above information by the processor. When outputting the above information, the processor outputs the above information to the transceiver so that it is transmitted by the transceiver. After the above information is output by the processor, it may also need to be processed in other ways before it reaches the transceiver. Similarly, the process of receiving information in the above method can be understood as the process of the processor receiving the input information. When the processor receives the input information, the transceiver receives the above information and inputs it into the processor. Furthermore, after the transceiver receives the above information, the above information may need to be processed in other ways before it is input into the processor.

[0301] See also Fig.10 , Fig.10 2 is a schematic diagram of the structure of another communication device provided in an embodiment of the present application. The communication device 20 may be the communication device 10, or a chip therein. Fig.10Only the main components of the communication device 20 are shown. In addition to the processor 1001, the communication device 20 may optionally further include a transceiver 1002, a memory 1003, or an input / output device (not shown).

[0302] The processor 1001 is mainly used to process the communication protocol and communication data, and to control the entire communication device, execute the software program, and process the data of the software program. The memory 1003 is mainly used to store the software program and data. The transceiver 1002 may include a control circuit and an antenna. The control circuit is mainly used to convert the baseband signal and the radio frequency signal and process the radio frequency signal. The antenna is mainly used to send and receive radio frequency signals in the form of electromagnetic waves. The input and output devices, such as a touch screen, a display screen, a keyboard, etc., are mainly used to receive data input by the user and output data to the user.

[0303] When the communication device is turned on, the processor 1001 can read the software program in the memory 1003, interpret and execute the instructions of the software program, and process the data of the software program. When data needs to be sent wirelessly, the processor 1001 performs baseband processing on the data to be sent, and outputs the baseband signal to the radio frequency circuit. The radio frequency circuit performs radio frequency processing on the baseband signal and then sends the radio frequency signal outward in the form of electromagnetic waves through the antenna. When data is sent to the communication device, the radio frequency circuit receives the radio frequency signal through the antenna, converts the radio frequency signal into a baseband signal, and outputs the baseband signal to the processor 1001. The processor 1001 converts the baseband signal into data and processes the data.

[0304] In another implementation, the RF circuit and antenna may be arranged independently of the processor performing baseband processing. For example, in a distributed scenario, the RF circuit and antenna may be arranged remotely from the communication device.

[0305] The transceiver 1002 may include a receiver and a transmitter, wherein the receiver is used to perform a receiving function (or operation) and the transmitter is used to perform a transmitting function (or operation), and the transceiver is used to communicate with other devices / apparatuses through a transmission medium.

[0306] The processor 1001 , the transceiver 1002 , and the memory 1003 may be connected via a communication bus.

[0307] Exemplarily, when the communication device 20 is used to execute the steps or methods or functions involved in the above-mentioned first communication device, the processor 1001 is used to send first information to the second communication device in the mobile communication network; the second communication device is deployed with a first sub-neural network model among the multiple sub-neural network models, and the first information is used to indicate a first operator, and the first operator is used to compress the service data transmitted by the second communication device to other communication devices, and the other communication devices are deployed with other sub-neural network models among the multiple sub-neural network models except the first sub-neural network model. Optionally, the transceiver 1002 can also be used to complete part of the functions of receiving or sending messages of the above-mentioned processor 1001.

[0308] Exemplarily, when the communication device 20 is used to execute the steps or methods or functions involved in the above-mentioned second communication device, the processor 1001 is used to receive the first information sent by the first communication device in the mobile communication network; the first information is used to indicate the first operator; the large model also includes a control node, and the control node is deployed on the first communication device; obtain the first business data, and input the first business data into the first sub-neural network model for processing to obtain the second business data; send the third business data to the third communication device in the mobile communication network; the third communication device is deployed with the second sub-neural network model in the large model, and the third business data is obtained by compressing the second business data based on the first operator, and the data volume of the third business data is less than the data volume of the second business data. Optionally, the transceiver 1002 can also be used to complete the function of partially receiving or sending messages of the above-mentioned processor 1001.

[0309] Exemplarily, when the communication device 20 is used to execute the steps or methods or functions involved in the third communication device, the processor 1001 is used to receive the second information sent by the first communication device in the mobile communication network; the second information is used to indicate the second operator; the large model also includes a control node, and the control node is deployed on the first communication device; receiving the third business data sent by the second communication device in the mobile communication network; the first sub-neural network model in the large model is deployed on the second communication device; inputting the fourth business data into the second sub-neural network model for processing to obtain the fifth business data; sending the fifth business data; the fourth business data is obtained by decompressing the third business data based on the second operator; the data volume of the third business data is less than the data volume of the fourth business data. Optionally, the transceiver 1002 can also be used to complete the function of partially receiving or sending messages of the above-mentioned processor 1001.

[0310] It is understandable that the specific description of the processor and the transceiver can also refer to the above Fig. 9 The introduction of the processing unit, the receiving unit and the sending unit in the device embodiment is not repeated here.

[0311] Optionally, the processor 1001 may include a transceiver for implementing the receiving and sending functions. For example, the transceiver may be a transceiver circuit, an interface, or an interface circuit. The transceiver circuit, interface, or interface circuit for implementing the receiving and sending functions may be separate or integrated. The above-mentioned transceiver circuit, interface, or interface circuit may be used for reading and writing code / data, or the above-mentioned transceiver circuit, interface, or interface circuit may be used for transmitting or delivering signals.

[0312] Optionally, the processor 1001 may store instructions, which may be computer programs. The computer programs run on the processor 1001, and the communication device 20 may execute the method described in the above method embodiment. The computer program may be fixed in the processor 1001, in which case the processor 1001 may be implemented by hardware.

[0313] In one implementation, the communication device 20 may include a circuit that can implement the functions of sending or receiving or communicating in the aforementioned method embodiment. The processor and transceiver described in the present application can be implemented in an integrated circuit (IC), an analog IC, a radio frequency integrated circuit (RFIC), a mixed signal IC, an application specific integrated circuit (ASIC), a printed circuit board (PCB), an electronic device, etc. The processor and transceiver can also be manufactured using various IC process technologies, such as complementary metal oxide semiconductor (CMOS), N-type metal oxide semiconductor (nMetal-oxide-semiconductor, NMOS), P-type metal oxide semiconductor (positive channel metal oxide semiconductor, PMOS), bipolar junction transistor (bipolar junction transistor, BJT), bipolar CMOS (BiCMOS), silicon germanium (SiGe), gallium arsenide (GaAs), etc.

[0314] It is understandable that the communication device shown in the embodiment of the present application may also have Fig.10More components and parts are not limited in the embodiments of the present application. The method executed by the processor and transceiver described above is only an example, and the specific steps executed by the processor and transceiver can refer to the introduction of the method embodiment above.

[0315] In another possible implementation, Fig. 9 In the communication device involved, the processing unit 102 may be one or more logic circuits; the sending unit 100 and the receiving unit 101 may be input / output interfaces, or may be called communication interfaces, or interface circuits, or interfaces, etc. Alternatively, the sending unit may be an output interface, and the receiving unit may be an input interface, and the sending unit and the receiving unit may be integrated into one unit, such as an input / output interface. Fig.11 , Fig.11 is a schematic diagram of the structure of another communication device provided in an embodiment of the present application. Fig.11 As shown, the communication device 30 includes a logic circuit 901 and an interface 902. That is, the processing unit 102 can be implemented by the logic circuit 901, and the sending unit 100 and the receiving unit 101 can be implemented by the interface 902. The logic circuit 901 can be a chip, a processing circuit, an integrated circuit or a system on chip (SoC) chip, etc., and the interface 902 can be a communication interface, an input and output interface, a pin, etc. For example, Fig.11 The communication device 30 is taken as an example of a chip, and the chip includes a logic circuit 901 and an interface 902 .

[0316] In the embodiment of the present application, the logic circuit and the interface may also be coupled to each other. The embodiment of the present application does not limit the specific connection method between the logic circuit and the interface.

[0317] The communication device 30 may be used to execute the method, function or step involved in the first communication device, the second communication device or the third communication device in the method embodiment.

[0318] Exemplarily, when the communication device 30 is used to execute the method, function or step involved in the above-mentioned first communication device, the interface 902 is used to send first information to the second communication device in the mobile communication network; the second communication device is deployed with a first sub-neural network model among the multiple sub-neural network models, and the first information is used to indicate a first operator, and the first operator is used to compress the service data transmitted by the second communication device to other communication devices, and the other communication devices are deployed with other sub-neural network models among the multiple sub-neural network models except the first sub-neural network model. Optionally, the logic circuit 901 can be used to generate messages sent by the transceiver 1002 or process messages received by the transceiver 1002.

[0319] It is understandable that the communication device 30 can also execute the methods, functions or steps involved in the above-mentioned second communication device or third communication device. The relevant description of the logic circuit 901 and the interface 902 can refer to the relevant description when the communication device 20 and the communication device 30 execute the methods, functions or steps involved in the first communication device, and will not be repeated here.

[0320] It is understood that the specific description of the logic circuit 901 and the interface 902 can also refer to the above Fig. 9 The description of the processing unit, the sending unit and the receiving unit involved in the device embodiment will not be repeated here.

[0321] It can be understood that the communication device shown in the embodiment of the present application can implement the method provided in the embodiment of the present application in the form of hardware, or can implement the method provided in the embodiment of the present application in the form of software, etc., and the embodiment of the present application is not limited to this.

[0322] for Fig.11 The specific implementation methods of the various embodiments shown can also refer to the above embodiments, which will not be described in detail here.

[0323] An embodiment of the present application also provides a mobile communication system, which includes a first communication device, a second communication device and a third communication device. The first communication device, the second communication device and the third communication device can be used to execute the method in any of the aforementioned embodiments.

[0324] In addition, the present application also provides a computer program, which is used to implement the operations and / or processing performed by the communication device 10, the communication device 20, and the communication device 30 in the method provided by the present application.

[0325] The present application also provides a readable storage medium, which stores computer code. When the computer code runs on a computer, the computer executes the operations and / or processing performed by the communication device 10, the communication device 20, and the communication device 30 in the method provided in the present application.

[0326] The present application also provides a computer program product, which includes a computer code or a computer program. When the computer code or the computer program runs on a computer, the operations and / or processing performed by the communication device 10, the communication device 20, and the communication device 30 in the method provided by the present application are executed.

[0327] The embodiment of the present application also provides a chip system, which includes a processor for supporting a device to implement the functions involved in any of the above embodiments, for example, generating or processing the information involved in the above communication method. In one possible design, the chip system also includes a memory, which is used to store program instructions and data necessary for the device. The chip system can be composed of a chip, or it can include a chip and other discrete devices.

[0328] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0329] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0330] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0331] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the technical effects of the solutions provided in the embodiments of the present application.

[0332] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0333] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or partly contributed to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a readable storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned readable storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., and other media that can store program codes.

[0334] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A communication method, characterized in that: A method for applying to a first communication device in a mobile communication network, wherein a control node of a large model is deployed on the first communication device, wherein the large model further comprises a plurality of sub-neural network models, and wherein the plurality of sub-neural network models are respectively deployed on different communication devices in the mobile communication network; and wherein the method comprises: Sending first information to a second communication device in the mobile communication network; a first sub-neural network model among the multiple sub-neural network models is deployed on the second communication device, the first information is used to indicate a first operator, the first operator is used to compress business data transmitted by the second communication device to other communication devices, and the other communication devices are deployed with other sub-neural network models among the multiple sub-neural network models except the first sub-neural network model.

2. The method according to claim 1, characterized in that The method further comprises: Sending second information to a third communication device in the mobile communication network; the third communication device is deployed with a second sub-neural network model among the multiple sub-neural network models, the second information is used to indicate a second operator, and the second operator is used to decompress the business data compressed by the first operator.

3. The method according to any one of claims 1 to 2, characterized in that: The first operator and / or the second operator are determined based on the model type and the model splitting point of the large model, or the first operator and / or the third operator are determined based on a training algorithm.

4. The method according to any one of claims 1 to 3, characterized in that The method further comprises: Sending third information to the second communication device; the third information includes an initial operator, and the initial operator is used to compress the service data transmitted by the second communication device to the other communication device; Sending fourth information to a third communication device in the mobile communication network; the fourth information includes a loss function, and the first operator is obtained based on training of the loss function and the initial operator.

5. The method according to any one of claims 1 to 4, characterized in that The method further comprises: Sending fifth information to the second communication device; the fifth information is used to indicate a third operator, the third operator is used to compress the service data transmitted by the second communication device to the other communication devices, and the compression rate of the third operator is different from the compression rate of the first operator.

6. The method according to claim 5, characterized in that The method further comprises: Sending sixth information to a third communication device in the mobile communication network; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the service data compressed by the third operator.

7. The method according to any one of claims 1 to 6, characterized in that The method further comprises: Sending seventh information to the second communication device; the seventh information includes first parameter information of the large model, and the seventh information is used to instruct the second communication device to load the first sub-neural network model of the large model based on the first parameter information; Sending an eighth information to a third communication device in the mobile communication network; the eighth information includes the second parameter information of the large model, and the eighth information is used to instruct the third communication device to load the second sub-neural network model of the large model based on the second parameter information.

8. A communication method, characterized in that: A second communication device in a mobile communication network is applied, wherein a first sub-neural network model in a large model is deployed on the second communication device, wherein the large model includes a plurality of sub-neural network models, and the plurality of sub-neural network models are respectively deployed on different communication devices in the mobile communication network; the method comprises: receiving first information sent by a first communication device in the mobile communication network; the first information is used to indicate a first operator; the large model further includes a control node, and the control node is deployed on the first communication device; Acquire first business data, and input the first business data into the first sub-neural network model for processing to obtain second business data; Sending third business data to a third communication device in the mobile communication network; the second sub-neural network model in the large model is deployed on the third communication device, and the third business data is obtained by compressing the second business data based on the first operator, and the data volume of the third business data is smaller than the data volume of the second business data.

9. The method according to claim 8, characterized in that The first operator is determined based on the model type and the model splitting point of the large model, or the first operator is determined by the second communication device and the third communication device based on a training algorithm.

10. The method according to any one of claims 8 to 9, characterized in that The method further comprises: Receive third information sent by the first communication device; the third information includes an initial operator, the third information is used to instruct the second communication device to train the initial operator, the initial operator is used to compress the service data transmitted by the second communication device to other communication devices, and the first operator is obtained based on the training of the initial operator.

11. The method according to any one of claims 8 to 10, characterized in that: The data volume of the third business data is smaller than that of the second business data, including: the second business data includes information of M neurons, the third business data includes information of N neurons, the N neurons are some of the M neurons, and N is a positive integer less than M.

12. The method according to any one of claims 8 to 11, characterized in that The method further comprises: Receive fifth information sent by the first communication device; the fifth information is used to indicate a third operator; the third operator is used to compress business data transmitted by the second communication device to other communication devices, the compression rate of the third operator is different from the compression rate of the first operator, and the change value of the data transmission rate between the second communication device and the third communication device is greater than or equal to a preset threshold.

13. The method according to any one of claims 8 to 12, characterized in that The method further comprises: receiving seventh information sent by the first communication device; the seventh information includes first parameter information of the large model; The first sub-neural network model of the large model is loaded based on the first parameter information.

14. The method according to any one of claims 8 to 13, characterized in that The first business data is a business request sent by a terminal device, and a response to the first business data is obtained based on processing of the third business data by the second sub-neural network model.

15. A communication method, characterized in that: A third communication device applied to a mobile communication network, wherein the second sub-neural network model in the large model is deployed on the third communication device, wherein the large model includes a plurality of sub-neural network models, and the plurality of sub-neural network models are respectively deployed on different communication devices in the mobile communication network; the method comprising: receiving second information sent by a first communication device in the mobile communication network; the second information is used to indicate a second operator; the large model further includes a control node, and the control node is deployed on the first communication device; Receiving third service data sent by a second communication device in the mobile communication network; the first sub-neural network model in the large model is deployed on the second communication device; The fourth business data is input into the second sub-neural network model for processing to obtain the fifth business data, and the fifth business data is sent; the fourth business data is obtained by decompressing the third business data based on the second operator; the data volume of the third business data is smaller than the data volume of the fourth business data.

16. The method according to claim 15, characterized in that The second operator is determined based on the model type and the model splitting point of the large model, or the second operator is determined by the first communication device and the second communication device based on a training algorithm.

17. The method according to any one of claims 15 to 16, characterized in that The method further comprises: Receive fourth information sent by the first communication device; the fourth information includes a loss function, the first operator of the second communication device is obtained based on the training of the loss function, and the first operator is used to compress the service data transmitted by the second communication device to other communication devices.

18. The method according to any one of claims 15 to 17, characterized in that The data volume of the third business data is smaller than that of the fourth business data, including: the third business data includes information of N neurons, the fourth business data includes information of K neurons, the N neurons are part of the K neurons, and N is a positive integer less than K.

19. The method according to any one of claims 15 to 18, characterized in that The method further comprises: Receive sixth information sent by the first communication device; the sixth information is used to indicate a fourth operator, and the fourth operator is used to decompress the business data compressed by the third operator.

20. The method according to any one of claims 15 to 19, characterized in that The method further comprises: receiving eighth information sent by the first communication device; the eighth information including second parameter information of the large model; The second sub-neural network model of the large model is loaded based on the second parameter information.

21. The method according to any one of claims 15 to 20, characterized in that The sending of the fifth service data comprises: The fifth service data is sent to the terminal device, where the fifth service data is a response to the service request initiated by the terminal device.

22. A communication device, characterized in that: The method comprises a module or a unit for executing the method according to any one of claims 1 to 21.

23. A communication device, characterized in that: The method comprises a processor, wherein the processor is used to implement the method according to any one of claims 1 to 21 through a logic circuit or executing a code instruction.

24. A computer-readable storage medium, characterized in that: The storage medium stores a computer program or an instruction, and when the computer program or the instruction is executed by the communication device, the method according to any one of claims 1 to 21 is implemented.

25. A computer program, characterized in that The computer program comprises instructions, and when the computer program is executed by a communication device, the method according to any one of claims 1 to 21 is implemented.

26. A chip system, characterized in that: The chip system includes a processor for supporting the device to implement the functions involved in the method as described in any one of claims 1-21.

27. A mobile communication system, characterized in that: The mobile communication system includes a first communication device, a second communication device and a third communication device, the first communication device is used to execute the method as described in any one of claims 1-7, the second communication device is used to execute the method as described in any one of claims 8-14, and the third communication device is used to execute the method as described in any one of claims 15-21.

Images