Communication method and communication apparatus

Abstract

A communication method and a communication apparatus, which are used for ensuring the end-to-end service latency in a terminal-network collaborative inference scenario. The method comprises: acquiring a computation duration of a first network-side model, wherein the first network-side model is a sub-model obtained after processing a first model and used for a network to perform inference, the first network-side model corresponds to a first terminal-side model, and the first terminal-side model is a sub-model obtained after processing the first model and used for a terminal to perform inference; and on the basis of the computation duration, sending a first model file of the first model or first association information, wherein the first association information is used for indicating a correspondence between bandwidth ranges and model groups, one model group comprises one network-side model and one terminal-side model, and the network-side model corresponds to the terminal-side model.

Description

Communication method and communication apparatus

[0001] Cross-reference to Related Applications

[0002] This application claims priority to the Chinese Patent Application No. 202411311308.8, filed on September 19, 2024, and entitled “A Communication Method and Communication Apparatus”, the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0003] The present application relates to the field of communication technology, and in particular to a communication method and a communication apparatus. BACKGROUND

[0004] With the development of artificial intelligence (AI) technology, AI models gradually evolve from traditional small-scale models (referred to as small models), such as convolution neutral network (CNN) and regression neutral network (RNN), to large-scale models (referred to as large models). At the same time, terminal applications based on large models are emerging, such as question and answer systems, text-to-image, text-to-video, and various natural language processing and multi-modal applications. In such applications, in order to improve user experience, it is usually necessary to reduce the end-to-end service latency.

[0005] When deploying a model, a network management system can determine whether to split the model and how to split and distribute the model based on the computing power of a terminal. For example, if the computing power of the terminal meets the requirements of model inference, the network management system determines to split the model and determines the splitting rate of the model based on the computing power of the terminal. Hereinafter, the complete model before splitting is referred to as model #1, and it is assumed that the splitting rate is 1:9, which means that 10% of model #1 (hereinafter referred to as model #2) is deployed to the terminal, and the remaining 90% of the model (hereinafter referred to as model #3) is deployed to the network side. Taking model #1 including 100 layers as an example, the first 10 layers of model #1 (i.e., model #2) can be deployed to the terminal, and the last 90 layers of the model (i.e., model #3) can be deployed to the network side. Model #2 and model #3 jointly constitute model #1. When subsequent inference using the model is needed, the terminal first performs inference based on the deployed model #2 to obtain an intermediate inference result, and transmits the intermediate inference result to the network side through the air interface, and the network side continues to perform inference based on the intermediate inference result using model #3 to obtain a final inference result.

[0006] The manner of joint inference by the terminal and the network side can be referred to as end-network collaborative inference. The end-network collaborative inference has the following advantages: first, since the computing resources of the terminal are also used for model inference, the system resource utilization rate can be maximized, which is beneficial to improving the model inference speed and accuracy; second, since the terminal participates in the model inference, it is possible to protect user privacy.

[0007] In the end-network collaborative inference scenario, the end-to-end service latency of user experience mainly depends on the air interface transmission latency of the inference result in the air interface transmission and the model inference latency. However, the air interface channel state and the computing task on the device for deploying the model are dynamically changed. Therefore, how to guarantee the end-to-end service latency is a problem that persons skilled in the art are studying. SUMMARY

[0008] Embodiments of the present application provide a communication method and a communication device for guaranteeing the end-to-end service latency in the end-network collaborative inference scenario.

[0009] In a first aspect, a communication method is provided, which can be executed by a first device. The first device can be a management service producer or a component in the management service producer. The management service producer can be deployed in a cross-domain management function unit or a domain management function unit. The management service producer can be a network device such as a network function (NF) or a server. Embodiments of the present application do not make any limitation. Taking the first device as the management service producer as an example, the method comprises the following steps: obtaining a computing duration of a first network side model, the first network side model being a sub-model of a first model for network inference after processing, the first network side model corresponding to a first terminal side model, the first terminal side model being a sub-model of the first model for terminal inference after processing; and sending a first model file or first association information of the first model according to the computing duration; wherein the first association information is used to indicate the corresponding relationship between a bandwidth range and a model group, one model group comprising one network side model and one terminal side model, and the one network side model and the one terminal side model corresponding to each other.

[0010] In embodiments of the present application, the model file of the first model is sent according to the computing duration (or also referred to as the edge inference duration) of the network side model, so as to redeploy the model; or the first association information can also be sent according to the computing duration of the network side model, so as to indicate the dynamic adjustment of the model group for inference according to the bandwidth range, to adjust the data amount of the inference result to be transmitted. Moreover, the air interface transmission latency is related to the data amount of the inference result to be transmitted. Therefore, by adjusting the data amount of the inference result to be transmitted, the air interface transmission latency can be adjusted, so as to guarantee the end-to-end service latency in the end-network collaborative inference scenario.

[0011] In a possible implementation, the first model file or the first association information of the first model is sent according to the calculation duration, including: the first association information is sent to the terminal and / or the first node according to the calculation duration, and the first node is a node in which the first network side model is deployed.

[0012] In the technical solution, in the end-to-network collaborative inference scenario, the terminal or the node (for example, the first node) in which the network side model is deployed can be used as an inference collaborative decision point to determine a model group for inference, and therefore the first association information can be sent to the terminal or the first node. The terminal or the first node can determine the model group for inference according to the first association information and notify the opposite end of the determination result, which helps to reduce the probability of using different model groups of sub-models for inference on the terminal side and the network side and improve the accuracy of model inference. Alternatively, the first association information can be sent to the terminal and the first node, which are used to determine the model group for inference, and therefore the terminal and the first node do not need to communicate, which helps to improve the model switching efficiency.

[0013] In a possible implementation, the first association information is sent to the terminal and / or the first node according to the calculation duration, including: determining a maximum communication delay according to the calculation duration and a service level agreement (SLA) requirement of the service corresponding to the first model; determining the first association information according to the maximum communication delay; and sending the first association information to the terminal and / or the first node.

[0014] In the technical solution, the maximum end-to-end delay can be determined according to the SLA requirement of the service corresponding to the first model. The maximum end-to-end delay can be decomposed into the calculation duration of the terminal side model, the calculation duration of the network side model, and the communication duration (for example, the communication delay), and the calculation duration of the terminal side model is relatively small and far less than the calculation duration of the network side model. Therefore, the maximum communication delay can be determined according to the maximum end-to-end delay and the calculation duration of the network side model, and the bandwidth range corresponding to different model groups can be determined according to the maximum communication delay, for example, the bandwidth range corresponding to different model groups is determined according to the data amount of the inference result of different model groups and the maximum communication delay, so that the terminal or the first node can select a suitable model group for inference according to the actual bandwidth to meet the SLA requirement of the service corresponding to the first model.

[0015] In a possible implementation, the first association information is used to indicate the correspondence between the bandwidth range and the model group, including: the first association information is used to indicate the correspondence between the bandwidth range and the collaboration level, and the collaboration level is associated with the model group.

[0016] In the technical solution, the first association information can also indicate a correspondence between the bandwidth range and the cooperation level, or can also indicate a correspondence between the bandwidth range and other information associated with the model group, and embodiments of the present application do not limit the content indicated by the first association information.

[0017] In a possible implementation, the first model file includes a first sub-model file corresponding to a second terminal-side model and a second sub-model file corresponding to a second network-side model, the second terminal-side model corresponding to the second network-side model, and the sending of the first model file or the first association information according to the calculation duration includes: sending first information including the first sub-model file to a terminal and sending second information including the second sub-model file to a first node according to the calculation duration, the first node being a node deploying the first network-side model.

[0018] In the technical solution, the first sub-model file of the terminal-side model can be sent to the terminal, and the second sub-model file of the network-side model can be sent to the first node, so as to redeploy the sub-models by the terminal and the first node, and the terminal and the first node can use the redeployed sub-models for inference after redeploying the sub-models, thereby guaranteeing the end-to-end service latency in the end-network collaborative inference scenario.

[0019] In a possible implementation, the method further includes: obtaining a third sub-model file corresponding to the first terminal-side model and a fourth sub-model file corresponding to the first network-side model; adjusting model parameters in the third sub-model file according to the calculation duration to obtain the first sub-model file; and adjusting model parameters in the fourth sub-model file according to the calculation duration to obtain the second sub-model file.

[0020] In the technical solution, the data amount of the inference result between the sub-models can be adjusted by modifying the model parameters, thereby guaranteeing the end-to-end service latency in the end-network collaborative inference scenario. For example, the codec and other parameters of the first model can be modified to adjust the feature compression ratio, model sparsity, and the like of the sub-models, so as to adjust the data amount of the inference result to be transmitted and guarantee the end-to-end service latency in the end-network collaborative inference scenario.

[0021] In a possible implementation, adjusting the model parameters in the third sub-model file according to the calculation duration and adjusting the model parameters in the fourth sub-model file according to the calculation duration include: determining an SLA of the first model corresponding service according to the calculation duration; and adjusting the model parameters in the third sub-model file according to the SLA and SLA requirement of the service, and adjusting the model parameters in the fourth sub-model file according to the SLA and SLA requirement of the service.

[0022] In the technical solution, the model parameters are adjusted according to the actual SLA and the SLA requirement, so that the reasoning accuracy of the model and the end-to-end service time delay can be ensured. For example, if the actual end-to-end time delay of the first model is less than the time delay corresponding to the SLA requirement, the model parameters can be adjusted according to the SLA requirement to increase the data amount of the reasoning results between the sub-models, so as to improve the reasoning accuracy of the model; if the actual end-to-end time delay of the first model is greater than the time delay corresponding to the SLA requirement, the model parameters can be adjusted according to the SLA requirement to reduce the data amount of the reasoning results between the sub-models, so as to ensure the end-to-end service time delay.

[0023] In a possible implementation, the third sub-model file corresponding to the first terminal-side model and the fourth sub-model file corresponding to the first network-side model are obtained by determining the third sub-model file and the fourth sub-model file according to the second model file of the first model.

[0024] In the technical solution, before the model parameters in the sub-model files are modified, the complete model file (i.e., the second model file) of the first model can be obtained first, and the second model file is divided into the sub-model file corresponding to the first terminal-side model and the sub-model file corresponding to the first network-side model.

[0025] In a possible implementation, the method further includes receiving third information, wherein the third information includes the second model file.

[0026] In the technical solution, the first device can not store the model file of the first model. Before the model parameters in the model file are modified, the complete model file (i.e., the second model file) of the first model can be obtained from other devices (e.g., devices of a third party), which helps to reduce the storage burden of the first device.

[0027] In a possible implementation, the method further includes receiving fourth information, wherein the fourth information includes the third sub-model file and the fourth sub-model file.

[0028] In the technical solution, before the model parameters in the sub-model files are modified, the third sub-model file and the fourth sub-model file can be directly obtained from other devices (e.g., devices of a third party), which does not require the first device to divide the second model file of the first model again, and helps to improve the efficiency of modifying the model parameters.

[0029] In a possible implementation, the method further includes obtaining the second model file of the first model, and determining the first sub-model file and the second sub-model file according to the calculation time length and the second model file.

[0030] In the technical solution, the data amount of inference results between the sub-models can be adjusted by re-dividing the model, thereby guaranteeing the end-to-end service delay in the end-network collaborative inference scenario.

[0031] In a possible implementation, the method further includes: receiving third information, the third information including the second model file.

[0032] In the technical solution, the first device can not store the model file of the first model, and the first device can obtain the complete model file (i.e., the second model file) of the first model from other devices (e.g., devices of a third party) before re-dividing the model file of the first model, thereby reducing the storage burden of the first device.

[0033] In a possible implementation, the method further includes: sending a first request to the terminal, the first request being used to request to delete a third sub-model file corresponding to the first terminal-side model; and sending a second request to the first node, the second request being used to request to delete a fourth sub-model file corresponding to the first network-side model.

[0034] In the technical solution, after instructing the terminal and the first node to re-deploy the model, the terminal and the first node can be further instructed to delete the sub-model file corresponding to the sub-model that is not used, thereby reducing the storage burden of the terminal and the first node.

[0035] In a possible implementation, before obtaining the calculation duration of the first network-side model, the method further includes: determining that a third request from a second device is received, the third request being used to request to optimize the SLA of the service corresponding to the first model, the third request including the identifier of the first model; or determining that the service level agreement (SLA) of the service corresponding to the first model does not meet the SLA requirement of the service.

[0036] In the technical solution, the first model file or the first associated information of the first model can be sent when the request to optimize the SLA of the service corresponding to the model is received, thereby not needing to monitor the calculation duration of the first network-side model in real time, and thereby reducing the operation burden of the first device; or the first model file or the first associated information of the first model can be sent when it is determined that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service, thereby guaranteeing the end-to-end service delay in the end-network collaborative inference scenario.

[0037] In a second aspect, a communication method is provided, which can be performed by a first device. The first device can be a management service producer or a component in the management service producer. The management service producer can be deployed in a cross-domain management function unit or a domain management function unit. The management service producer can be a network device such as a network function (NF) or a server. The embodiments of the present application are not limited thereto. Taking the first device as the management service producer as an example, the method comprises: receiving a third request from a second device, the third request being used to request optimization of SLA of a service corresponding to a first model, the third request comprising an identifier of the first model; or determining that SLA of the service corresponding to the first model does not meet SLA requirements of the service; determining a second node with the largest remaining resources in a plurality of nodes, and deploying a first network-side model in the second node, the first network-side model being a sub-model of the first model processed for network inference.

[0038] In the embodiments of the present application, when the first device receives a request for optimization of SLA of a service corresponding to a model, or determines that SLA of the service corresponding to the first model does not meet SLA requirements of the service, the first network-side model of the first model is deployed on a node with the largest remaining resources, which helps to reduce the calculation delay (or also referred to as inference delay) of the first network-side model, thereby guaranteeing end-to-end service delay in an end-network collaborative inference scenario.

[0039] In a possible implementation, deploying the first network-side model in the second node comprises: sending a fourth request to the second node, the fourth request being used to request deployment of the first network-side model, the fourth request comprising a model file corresponding to the first network-side model.

[0040] In the technical solution, when it is determined that the deployment node of the network-side model needs to be replaced, a request can be sent to the target node (the second node) to deploy the network-side model, which meets the existing protocol specification.

[0041] In a possible implementation, the method further comprises: sending a fifth request to the first node, the fifth request being used to request deletion of a model file corresponding to the first network-side model, the fifth request comprising an identifier of the first model.

[0042] In the technical solution, after instructing the second node to deploy the network-side model, the first node can also be instructed to delete a sub-model file corresponding to the network-side model of the first model, which reduces the running and storage burden of the first node. In addition, it can also reduce the probability of causing inference conflicts when the first node and the second node both perform model inference.

[0043] In a third aspect, a communication apparatus is provided. The communication apparatus can be the first apparatus of the first aspect or the second aspect. The communication apparatus has the functions of the first apparatus. For example, the communication apparatus has the functions of the first apparatus of the first aspect or the second aspect. For example, the communication apparatus includes modules or units or means corresponding to the operations of the first apparatus of the first aspect or the second aspect. The modules or units or means can be implemented by software or by hardware or by a combination of software and hardware.

[0044] In another optional implementation, the communication apparatus includes a processing unit (also sometimes referred to as a processing module) and a transceiving unit (also sometimes referred to as a transceiving module). The transceiving unit can implement the sending function and the receiving function. When the transceiving unit implements the sending function, it can be referred to as a sending unit (also sometimes referred to as a sending module). When the transceiving unit implements the receiving function, it can be referred to as a receiving unit (also sometimes referred to as a receiving module). The sending unit and the receiving unit can be the same functional module, which is referred to as a transceiving unit. The transceiving unit can implement the sending function and the receiving function. Alternatively, the sending unit and the receiving unit can be different functional modules, and the transceiving unit is a general term for these functional modules.

[0045] In an optional implementation, the processing unit is configured to obtain a calculation duration of a first network-side model. The first network-side model is a sub-model of the first model after processing, which is used for inference by a network. The first network-side model corresponds to a first terminal-side model. The first terminal-side model is a sub-model of the first model after processing, which is used for inference by a terminal.

[0046] The transceiving unit (or the sending unit) is configured to send a first model file of the first model or first association information according to the calculation duration. The first association information is used to indicate a correspondence between a bandwidth range and a model group. One model group includes one network-side model and one terminal-side model. The one network-side model and the one terminal-side model correspond to each other.

[0047] In yet another optional implementation, the transceiving unit (or the receiving unit) is configured to receive a third request from a second apparatus. The third request is used to request optimization of an SLA of a service corresponding to the first model. The third request includes an identifier of the first model.

[0048] The processing unit is configured to determine a second node with the largest remaining resource among the plurality of nodes and deploy the first network-side model on the second node when the transceiving unit receives the third request or the processing unit determines that the SLA of the first model corresponding to the service does not satisfy the SLA requirement corresponding to the service, the first network-side model being a sub-model of the first model processed for network inference.

[0049] In an optional implementation, the communication apparatus further includes a storage unit (also referred to as a storage module), and the processing unit is configured to be coupled with the storage unit and execute programs or instructions in the storage unit to enable the communication apparatus to perform the functions of the first apparatus in the first aspect or the second aspect.

[0050] In a fourth aspect, a communication apparatus is provided, which includes a memory and one or more processors. The memory is configured to store part or all of necessary computer programs or instructions for implementing the functions in the first aspect or the second aspect. The one or more processors are configured to execute the computer programs or instructions, which, when executed, cause the communication apparatus to implement the method in any possible design or implementation manner of the first aspect or the second aspect.

[0051] In a possible design, the communication apparatus can further include an interface circuit, and the processor is configured to communicate with other apparatuses or components through the interface circuit.

[0052] In a possible design, the communication apparatus can further include the memory.

[0053] The communication apparatus can be a network device, a communication module in a network device, or a chip responsible for communication functions in a network device, such as a modem chip (also referred to as a baseband chip) or a system on chip (SoC) or a system in package (SIP) chip including a modem module. Optionally, the network device can implement the method in any possible design or implementation manner of the first aspect, or can implement the method in any possible design or implementation manner of the second aspect.

[0054] In a fifth aspect, a computer readable storage medium is provided, which is configured to store computer programs or instructions, which, when executed, cause the method performed by the first apparatus, the terminal, the first node or the second apparatus in the aspects to be implemented.

[0055] In a sixth aspect, a computer program product including instructions is provided, which, when executed on a computer, causes the method in the aspects to be implemented.

[0056] In a seventh aspect, a chip system is provided, comprising a processor and an interface, the processor is configured to invoke and run instructions from the interface, so that the chip system implements the method of the above aspects.

[0057] The technical effects achieved by any of the above third to seventh aspects can be described with reference to the technical effects achieved by any of the possible implementation manners of the first or second aspect, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0058] FIG. 1 is a communication system to which embodiments of the present application are applied;

[0059] FIG. 2 is a schematic diagram of a system architecture provided by embodiments of the present application;

[0060] FIG. 3 is a schematic diagram of a flow of a first communication method provided by embodiments of the present application;

[0061] FIG. 4 is a schematic diagram of a flow of an example of the communication method shown in FIG. 3 provided by embodiments of the present application;

[0062] FIG. 5 is a schematic diagram of a flow of another example of the communication method shown in FIG. 3 provided by embodiments of the present application;

[0063] FIG. 6 is a schematic diagram of a flow of a second communication method provided by embodiments of the present application;

[0064] FIG. 7 is a schematic diagram of an apparatus provided by embodiments of the present application;

[0065] FIG. 8 is a schematic diagram of another apparatus provided by embodiments of the present application. DETAILED DESCRIPTION

[0066] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be described in further detail below with reference to the accompanying drawings. The specific operation methods in the method embodiments can also be applied to the apparatus embodiments or system embodiments.

[0067] The terms "system" and "network" in the embodiments of the present application can be used interchangeably. "Multiple" means two or more, and therefore, "multiple" can also be understood as "at least two" in the embodiments of the present application. "At least one" can be understood as one or more, for example, as one, two or more. For example, including at least one means including one, two or more, and does not limit which ones are included. For example, including at least one of A, B and C means that A, B, C, A and B, A and C, B and C, or A and B and C can be included. Similarly, the understanding of "at least one" and the like descriptions is similar. "At least one of the following" or the like means any combination of these items, including any combination of single item or multiple items. For example, "at least one of A, B and C" includes A, B, C, AB, AC, BC or ABC. "And / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. In addition, the character " / ", unless otherwise specified, generally indicates that the associated objects before and after are in an "or" relationship.

[0068] Unless otherwise specified, the ordinal numbers "first", "second", "third", "fourth" and the like mentioned in the embodiments of the present application are used to distinguish a plurality of objects, and are not used to limit the order, time sequence, priority or importance of the plurality of objects, and the descriptions of "first", "second", "third" and "fourth" do not necessarily mean that the objects are different.

[0069] Embodiments of the present application can be applied to various mobile communication systems, such as a new radio (NR) system, a long term evolution (LTE) system, an advanced long term evolution (LTE-A) system, a future communication system, and other communication systems, without limitation. For example, refer to FIG. 1 for a communication system to which embodiments of the present application are applicable. As shown in FIG. 1, the communication system 10 includes a radio access network (RAN) 100 and a core network (CN) 200. Optionally, the communication system further includes an Internet 300. The RAN 100 includes at least one RAN node (e.g., 110a and 110b in FIG. 1, collectively referred to as 110) and at least one terminal (e.g., 120a-120j in FIG. 1, collectively referred to as 120). The RAN 100 can further include other RAN nodes, such as a wireless relay device and / or a wireless backhaul device (not shown in FIG. 1), etc. The terminal 120 is connected to the RAN node 110 in a wireless manner. The RAN node 110 is connected to the core network 200 in a wireless or wired manner. The core network device in the core network 200 and the RAN node 110 in the RAN 100 can be different physical devices respectively, or can be the same physical device integrated with the logical functions of the core network and the radio access network.

[0070] The RAN 100 can be a 3rd generation partnership project (3GPP) related cellular system, such as a fourth generation (4G), a fifth generation (5G) mobile communication system, or a future-oriented evolution system. The RAN 100 can also be an open radio access network (O-RAN or ORAN), a cloud radio access network (CRAN), or a wireless fidelity (WiFi) system. The RAN 100 can also be a communication system in which two or more of the above systems are fused.

[0071] The RAN node 110, which can also be referred to as an access network device, a RAN entity, or an access node, etc., forms part of the communication system, and is configured to facilitate the wireless access by the terminals. The RAN nodes 110 in the communication system 10 can be of the same type or can be of different types. In some scenarios, the roles of the RAN node 110 and the terminal 120 are relative, e.g., the network element 120i in Figure 1 can be a helicopter or a drone, which can be configured to move as a mobile base station, and for a terminal 120j accessing the RAN 100 via the network element 120i, the network element 120i is a base station; but for the base station 110a, the network element 120i is a terminal. The RAN nodes 110 and the terminals 120 are sometimes referred to as communication devices, e.g., the network elements 110a and 110b in Figure 1 can be understood as communication devices with base station functionalities, and the network elements 120a-120j can be understood as communication devices with terminal functionalities.

[0072] In a possible scenario, the RAN node can be a base station, an evolved NodeB (eNodeB), an access point (AP), a transmission reception point (TRP), a next generation NodeB (gNB), a base station in a future mobile communication system, or an access node in a WiFi system, etc. The RAN node can be a macro base station (e.g., 110a in Figure 1), a micro base station or an indoor station (e.g., 110b in Figure 1), a relay node or a donor node, or a wireless controller in a CRAN scenario. Optionally, the RAN node can also be a server, a wearable device, a vehicle or a vehicle-mounted device, etc. For example, the access network device in vehicle to everything (V2X) technology can be a road side unit (RSU). All or part of the functions of the RAN node in this application can also be implemented by software functions running on hardware, or by virtualized functions instantiated on a platform (e.g., a cloud platform). The RAN node can also be provided with a communication module, circuit or chip for performing corresponding communication functions, and program instructions for performing corresponding communication functions. The RAN node in this application can also be a logical node, a logical module or software that can implement all or part of the functions of the RAN node.

[0073] In another possible scenario, a terminal is assisted by multiple RAN nodes to implement wireless access, and different RAN nodes respectively implement part of functions of a base station. For example, a RAN node can be a central unit (CU), a distributed unit (DU), a CU-control plane (CP), a CU-user plane (UP), or a radio unit (RU), etc. The CU and the DU can be separately arranged, or can also be included in the same network element, for example, in a baseband unit (BBU). The RU can be included in a radio frequency device or a radio frequency unit, for example, included in a remote radio unit (RRU), an active antenna processing unit (AAU), or a remote radio head (RRH).

[0074] In different systems, the CU (or CU-CP and CU-UP), DU or RU can also have different names, but those skilled in the art can understand their meanings. For example, in an ORAN system, the CU can also be referred to as an O-CU (open CU), the DU can also be referred to as an O-DU, the CU-CP can also be referred to as an O-CU-CP, the CU-UP can also be referred to as an O-CU-UP, and the RU can also be referred to as an O-RU. For the convenience of description, the CU, CU-CP, CU-UP, DU and RU are taken as examples for description in this application. Any one of the CU (or CU-CP, CU-UP), DU and RU in this application can be implemented by a software module, a hardware module, or a combination of a software module and a hardware module.

[0075] A terminal can be a device or module with corresponding communication functions and can access the above-mentioned communication system. The terminal can also be referred to as a terminal device, user equipment (UE), mobile station, mobile terminal, etc. The terminal can be widely used in various scenarios, such as device-to-device (D2D) communication, vehicle-to-everything (V2X) communication, machine-type communication (MTC), internet of things (IOT), virtual reality, augmented reality, industrial control, autonomous driving, remote medical treatment, smart power grid, smart furniture, smart office, smart wear, smart transportation, smart city, etc. The terminal can be a mobile phone, tablet computer, computer with wireless transceiver function, wearable device, vehicle, unmanned aerial vehicle, helicopter, airplane, ship, robot, mechanical arm, smart home device, transport vehicle with wireless communication function, communication module, etc. Embodiments of the present application do not limit the device form of the terminal. The terminal usually has a communication module, circuit or chip for performing corresponding communication functions. The terminal also has program instructions for performing corresponding communication functions.

[0076] FIG. 2 is a schematic diagram of a system architecture provided by embodiments of the present application, which can also be referred to as a network management system architecture, etc. The system architecture includes a management service consumer, a management service provider and a network element. The management service provider includes a cross-domain management function unit and a domain management function unit. Each unit will be briefly introduced as follows.

[0077] (1) Management service consumer.

[0078] The management service consumer is used to request and invoke management services. For example, the management service consumer can initiate a management service invocation request to a management domain (such as a cross-domain management unit or a single-domain management unit). An example of the management service consumer is a 3rd party entity, such as a vertical industry management system, which can initiate a management service invocation request to a management domain (cross-domain management system or single-domain management system).

[0079] (2) Cross-domain management function unit.

[0080] The cross-domain management function unit, which can also be referred to as a network management function (NMF) or a network management system (NMS), is taken as an example of the NMS in the following embodiments. The cross-domain management function unit is responsible for the operation, management, and maintenance functions of the network. One cross-domain management function unit can manage one or more domain management function units, such as a domain management function unit managing a RAN domain and a domain management function unit managing a CN domain.

[0081] (3) Domain management function unit.

[0082] The domain management function unit, which can also be referred to as an element management system (EMS), is taken as an example of the EMS in the following embodiments. The EMS includes, for example, a subnetwork management function (NMF), a network element / function management function, an EMS of a RAN domain, or an EMS of a core network domain. One EMS can manage one or more network elements, such as an EMS of a RAN domain managing access network devices, and an EMS of a CN domain managing a network data analytics function (NWDAF) network element, a user plane function (UPF) network element, a session management function (SMF), an access and mobility management function (AMF), and the like.

[0083] (4) Network element.

[0084] A network element is an entity for providing network services, including a radio access network network element, a core network network element (not shown in the figure), a transport network network element (not shown in the figure), and the like. For example, the core network network element can include, but is not limited to, an access and mobility management function (AMF) entity, a session management function (SMF) entity, a policy control function (PCF) entity, a user plane function (UPF) entity, a network data analysis function (NWDAF) entity, a network repository function (NRF) entity, a gateway, and the like.

[0085] In addition, in the ORAN, a service management and orchestration (SMO) system is also included, and the role of the SMO system in the network architecture is similar to that of the NMS, and the SMO system is responsible for the operation, management and maintenance of network services and orchestration functions of each domain, for example, can manage network elements (such as access network devices, NWDAF network elements, etc.) of each domain.

[0086] Data or model information and the like in the network can be opened to an over the top (OTT) server of a terminal manufacturer by the NMS or the SMO system, and the data or model information and the like are sent to the terminal by the OTT server.

[0087] With the development of AI technology, AI models gradually evolve from traditional small-scale models (referred to as small models), such as CNN and RNN, to large-scale models (referred to as large models). At the same time, terminal applications based on large models are emerging, such as question and answer systems, text-to-image, text-to-video, and various natural language processing and multi-modal applications. In such applications, in order to improve user experience, it is usually necessary to reduce the end-to-end service latency.

[0088] When the model is deployed, the network management system can determine whether to split the model and how to split and perform distributed deployment based on the computing capability of the terminal. For example, if the computing capability of the terminal meets the requirements of model inference, the network management system determines to split the model, and determines the splitting rate of the model based on the computing capability of the terminal. The complete model before splitting is referred to as model #1 below, and it is assumed that the splitting rate is 1:9, that is, 10% of model #1 (referred to as model #2 below) is deployed to the terminal, and the remaining 90% of the model (referred to as model #3 below) is deployed to the network side. Taking an example in which model #1 includes 100 layers, the first 10 layers of model #1 (that is, model #2) can be deployed to the terminal, and the last 90 layers of the model (that is, model #3) can be deployed to the network side. Model #2 and model #3 jointly constitute model #1. When subsequent inference using the model is required, the terminal first performs inference based on the deployed model #2 to obtain an intermediate inference result, and transmits the intermediate inference result to the network side through the air interface. The network side continues to perform inference based on the intermediate inference result using model #3 to obtain a final inference result.

[0089] The joint inference by the terminal and the network side can be referred to as end-network collaborative inference. The end-network collaborative inference has the following advantages: first, since the computing resources of the terminal are also used for model inference, the system resource utilization rate can be maximized, which is beneficial to improving the model inference speed and accuracy; second, since the terminal participates in the model inference, it is possible to protect user privacy.

[0090] In the scenario of end-network collaborative inference, the end-to-end service delay of user experience mainly depends on the air interface transmission delay of transmitting the intermediate inference result through the air interface and the model inference delay. However, the air channel state and the computing task on the device used to deploy the model are dynamically changing. Therefore, how to guarantee the end-to-end service delay is a problem that persons skilled in the art are studying.

[0091] In view of this, an embodiment of the present application provides a communication method, in which a first device (for example, an NMS) sends a model file of a first model according to the computing time length (or also referred to as edge inference time length) of a network side model to redeploy the model; or also sends first association information according to the computing time length of the network side model to indicate dynamically adjusting a model group used for inference according to a bandwidth range, so as to adjust the data amount of the inference result to be transmitted. Moreover, the air interface transmission delay is related to the data amount of the inference result to be transmitted. Therefore, by adjusting the data amount of the inference result to be transmitted, the air interface transmission delay can be adjusted, so as to guarantee the end-to-end service delay in the scenario of end-network collaborative inference.

[0092] The communication method and device are described below with reference to the accompanying drawings. It can be understood that the first device, the terminal, and the first node are taken as an example of the execution subject of the interaction in the present application, but the present application does not limit the execution subject of the interaction. For example, the method performed by the first device in the present application can also be implemented by a module (such as a circuit, a chip, or a chip system) in the first device or a logical node, a logical module, or software capable of implementing all or part of the function of the first device. The method performed by the terminal in the present application can also be implemented by a module (such as a circuit, a chip, or a chip system) in the terminal or a logical node, a logical module, or software capable of implementing all or part of the function of the terminal. The method performed by the first node in the present application can also be implemented by a module (such as a circuit, a chip, or a chip system) in the first node or a logical node, a logical module, or software capable of implementing all or part of the function of the first node.

[0093] The communication method provided by the embodiments of the present application is described in detail below with reference to the system architecture shown in FIG. 2.

[0094] In the present application, the first device can be, for example, an EMS, an NMS, or an SMO system shown in FIG. 2, or a device in the EMS, a device in the NMS, or a device in the SMO system, and the first node can be, for example, a network element (such as an access network element) shown in FIG. 2, or a server deployed by the network element shown in FIG. 2, or a building base band unit (BBU) inside a piece of computing power board in the server.

[0095] FIG. 3 is a flowchart of a first communication method provided by an embodiment of the present application. The method includes the following steps:

[0096] S301: The first device obtains the calculation time length of the first network-side model.

[0097] The joint inference mode of the terminal and the network side is called end-network collaborative inference. In the scenario of end-network collaborative inference, the model is distributed in the terminal and the network. The sub-model deployed in the terminal can also be understood as a sub-model for the terminal to perform inference, which can be referred to as a terminal-side model. The sub-model deployed in the network side can also be understood as a sub-model for the network to perform inference, which can be referred to as a network-side model.

[0098] The first network-side model is a sub-model obtained by processing the first model for the network to perform inference. The first network-side model corresponds to the first terminal-side model, and the first terminal-side model is a sub-model obtained by processing the first model for the terminal to perform inference.

[0099] The first device can obtain the calculation duration of the first network-side model in a manner such as: the first device can send a request to a node (e.g., the first node) that deploys the first network-side model, to request the calculation duration of the first network-side model. The calculation duration can be, for example, an average duration of the first network-side model in the last N times of performing an inference task, or can be a duration of the first network-side model in the last time of performing an inference task, and the present application does not limit the calculation duration. Optionally, the request can further include an identifier of the first model, which can be an application identifier (APP ID) or a service ID.

[0100] Optionally, before obtaining the calculation duration of the first network-side model, the first device can further determine whether a request (e.g., a third request described below) for optimizing the SLA of the service corresponding to the first model is received from a second device. The third request includes an identifier of the first model, which can be an APP ID or a service ID. For example, if the first device receives the third request, it indicates that the application developer wants to optimize (e.g., shorten) the SLA of the service corresponding to the first model, and the first device can obtain the calculation duration of the first network model. If the first device does not receive the third request, it indicates that the application developer can not have the demand to optimize the SLA of the service corresponding to the first model, and the first device can not obtain the calculation duration of the first network-side model. The second device is, for example, a third-party entity, and the third party is, for example, an entity of the application developer of the service corresponding to the first model. The SLA includes an end-to-end service delay, and therefore the third request can also be understood as a request for the end-to-end service delay of the service corresponding to the first model. In the present application, the SLA and the end-to-end service delay can be used interchangeably.

[0101] Alternatively, before obtaining the calculation duration of the first network-side model, the first device can further determine whether the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model. For example, if the first device determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, it indicates that the end-to-end service delay of the service corresponding to the first model can be long, and the first device can obtain the calculation duration of the first network-side model. If the first device determines that the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model, it indicates that the end-to-end service delay of the service corresponding to the first model is short, and the first device can not obtain the calculation duration of the first network-side model.

[0102] The first device determines whether the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model in the following manner, for example. The first device determines according to the calculation time length of the first model (including the calculation time length of the first terminal-side model and the calculation time length of the first network-side model), the communication time length, and the end-to-end service time delay indicated by the SLA requirement of the service corresponding to the first model.

[0103] Alternatively, the first device determines whether the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model in the following manner. For example, if the first device receives indication information from another device indicating that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, it can be determined that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model. If the first device does not receive indication information from the other device indicating that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, it can be determined that the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model. The other device may, for example, be indicated by at least one of a terminal, a first node, or an entity of a third party (for example, the second device described above). The embodiments of the present application do not limit the manner in which the first device determines whether the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model.

[0104] It can be understood that the first device receives the third request or determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model. The calculation time length of the first network-side model is only two examples. In a specific implementation process, the first device can also periodically obtain the calculation time length of the first network-side model for periodically adjusting the end-to-end service time delay of the service corresponding to the first model. The embodiments of the present application do not limit the conditions or timing of the first device obtaining the calculation time length of the first network-side model.

[0105] The end-to-end service time delay of the service corresponding to the first model can be divided into a communication time delay and a calculation time delay of the first model, wherein the calculation time delay of the first model includes a calculation time length of the first terminal-side model and a calculation time length of the first network-side model. The calculation time length of the first terminal-side model is the time length required for model input data to input the first terminal-side model to obtain a reasoning result by model inference. The communication time length is the time length required for the reasoning result to be sent to the first node. The calculation time length of the first network-side model is the time length required for the reasoning result to be input to the first network-side model to obtain a reasoning result by model inference. In the embodiments of the present application, the time length can also be understood as the time delay, that is, the time length and the time delay can be used alternately.

[0106] S302: The first device sends a first model file or first association information of the first model according to the first time length. Correspondingly, the terminal and / or the first node receives the first model file or the first association information of the first model.

[0107] In the scenario of end-to-network collaborative reasoning, the first model can include one or more sets of distributed deployment schemes, each set of distributed deployment schemes corresponding to a model group, and each model group corresponding to a model processing manner. For example, the first model processed by a model processing manner can obtain the aforementioned first terminal-side model and first network-side model, and processed by another model processing manner can obtain terminal-side model 1 and network-side model 1, which correspond to each other. That is, each model group includes a terminal-side model and a network-side model, and the terminal-side model and the first network-side model correspond to each other.

[0108] Optionally, when the first model includes multiple sets of distributed deployment schemes, the model segmentation ratios of the multiple sets of distributed deployment schemes can be the same, that is, the first model can be divided into multiple model groups according to the same model segmentation ratio, and each model group includes a terminal-side model and a network-side model. The model segmentation ratio is used to indicate the ratio of dividing the first model into a terminal-side model and a network-side model. For example, the model segmentation ratio is 1:9, which means that the first 10% of the layers of the first model are divided into a terminal-side model used by the terminal for model reasoning, and the last 90% of the layers of the first model are divided into a network-side model used by the network (for example, the first node) for model reasoning. The model segmentation ratio is related to the computing power and / or memory size of the terminal, for example, the stronger the computing power of the terminal, the more layers of the terminal-side model are divided to the terminal.

[0109] The multiple model groups correspond to the same model segmentation ratio, but the data amount of the inference result of the terminal-side model in different model groups is different. For example, the compression rates of the output layer matrices of different model groups are different. In the case that the data amount of the model input data is the same, the higher the compression rate of the output layer matrix of the terminal-side model, the less the data amount of the inference result of the terminal-side model; on the contrary, the lower the compression rate of the output layer matrix of the terminal-side model, the more the data amount of the inference result of the terminal-side model. The model input data can be understood as the data input into the first terminal-side model, and the inference result of the terminal-side model can be understood as the output data of the terminal-side model. In the embodiments of the present application, the inference result of the terminal-side model and the output data of the terminal-side model can be used alternately.

[0110] Taking the first model as an example, the model #1 is divided into three model groups according to the same model segmentation ratio and different model processing manners, the first model group (for example, model group 1) includes the terminal side model #1_1 and the network side model #1_1, the second model group (for example, model group 2) includes the terminal side model #1_2 and the network side model #1_2, and the third model group (for example, model group 3) includes the terminal side model #1_3 and the network side model #1_3. Assuming that the compression rate of the output layer matrix of the terminal side model #1_1 is the highest, the compression rate of the output layer matrix of the terminal side model #1_2 is the second, and the compression rate of the output layer matrix of the terminal side model #1_3 is the lowest, and the data amount of the model input data is the same, the data amount of the output data of the terminal side model #1_1 is the least, for example, 100 kilobytes (KB); the data amount of the output data of the terminal side model #1_2 is the second, for example, 500 KB; and the data amount of the output data of the terminal side model #1_3 is the most, for example, 1 megabytes (MB).

[0111] In addition, the network side model #1_1, the network side model #1_2 and the network side model #1_3 are also different, the network side model #1_1 corresponds to the terminal side model #1_1, the input data of the network side model #1_1 is the output data of the terminal side model #1_1, the network side model #1_2 corresponds to the terminal side model #1_2, the input data of the network side model #1_2 is the output data of the terminal side model #1_2, and the network side model #1_3 corresponds to the terminal side model #1_3, the input data of the network side model #1_3 is the output data of the terminal side model #1_3. That is, in the case that the data amount of the model input data is the same, the transmission data amount corresponding to different model groups is different, and the transmission data amount is the data amount of the data sent by the terminal to the first node (the output data of the terminal side model).

[0112] Optionally, when the first model includes multiple sets of distributed deployment schemes, that is, multiple terminal side models corresponding to the first model are deployed on the terminal, and multiple network side models corresponding to the first model are deployed on the first node, the terminal and the first node will cooperate to determine which model group corresponding to the set of distributed deployment schemes is used for model inference.

[0113] Optionally, the terminal and the first node can determine which model group to use for model inference according to the correspondence between the bandwidth range and the model group. For example, the terminal or the first node can be taken as an inference collaborative decision point, the inference collaborative decision point stores the correspondence between the bandwidth range and the model group, the inference collaborative decision point can perceive the air interface bandwidth (which can also be referred to as uplink transmission rate), determine which model group to use for model inference according to the air interface bandwidth and the correspondence between the bandwidth range and the model group, and send the determination result to the opposite end. For example, if the terminal is taken as the collaborative inference decision point, the terminal can perceive the air interface bandwidth, determine which model group to use for model inference according to the air interface bandwidth and the correspondence between the bandwidth range and the model group, and send the determination result to the first node. If the first node is taken as the collaborative inference decision point, the first node can perceive the bandwidth range, determine which model group to use for model inference according to the bandwidth range, and send the determination result to the terminal. It can be understood that when the air interface bandwidth is perceived by the collaborative inference decision point to determine which model group to use for model inference according to the correspondence between the bandwidth range and the model group, the collaborative inference decision point can also be other devices, for example, it can be other possible computing task control nodes, the other devices can perceive the air interface bandwidth, and the other devices can communicate with at least one of the terminal and the first node to indicate the model group for model inference, and the embodiments of the present application do not limit the collaborative inference decision point.

[0114] Taking the above three model groups as an example, the correspondence between the bandwidth range and the model group is shown in Table 1.

[0115] Table 1

[0116] In Table 1, bps is bits per second.

[0117] For example, taking the inference collaborative decision point as the first node as an example, if the first node perceives that the air interface bandwidth is 15bps, the first node can determine to use the model group 1 for model inference; if the first node perceives that the air interface bandwidth is 30bps, the first node can determine to use the model group 2 for model inference; if the first node perceives that the air interface bandwidth is 80bps, the first node can determine to use the model group 3 for model inference. When it is determined to use the model group 3 for model inference, the model group 3 can be understood as the currently (or currently) used model group, and the model group 1 and the model group 2 can be understood as the model groups to be used (or not used).

[0118] It can be understood that the greater the compression rate of the terminal side model, the lower the model inference accuracy, and vice versa. Therefore, when the air interface bandwidth is large, the model group corresponding to the terminal side model with a smaller compression rate can be preferentially selected to ensure the accuracy of model inference.

[0119] After the first device obtains the calculation duration of the first network side model, the first device can adjust the terminal side model used for terminal inference and the network side model used for network inference according to the calculation duration of the first network side model.

[0120] Optionally, the first device obtaining the calculation duration of the first network side model as described in S301 can be obtained when the first device receives the third request from the second device, or can be obtained when the first device determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, or can be obtained periodically by the first device.

[0121] If the first device obtains the first duration when receiving the third request from the second device, or determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, it indicates that the end-to-end service delay of the service corresponding to the first model needs to be shortened, and therefore the first device can adjust the terminal side model used for terminal inference and the network side model used for network inference according to the calculation duration of the first network side model, so that the data amount of the inference result of the adjusted terminal side model used for terminal inference is less than the data amount of the inference result of the terminal side model used for terminal inference before adjustment, thereby shortening the transmission delay of the inference result of the terminal side model used for terminal inference, and guaranteeing the SLA of the service corresponding to the first model.

[0122] If the first device acquires the first time length periodically, it indicates that the end-to-end service delay of the first model corresponding service can be shortened, or the end-to-end service delay of the first model corresponding service can also be lengthened. For example, if the actual end-to-end delay of the first model corresponding service is much smaller than the end-to-end service delay indicated by the SLA requirement of the first model corresponding service, the compression rate of the terminal side model can be reduced (that is, the end-to-end service delay of the first model corresponding service is lengthened) to increase the model inference accuracy. That is, on the basis of being able to guarantee the end-to-end service delay, a little end-to-end service delay can be sacrificed to improve the model inference accuracy. The related description of shortening the end-to-end service delay of the first model corresponding service can refer to the related description of shortening the end-to-end service delay of the first model corresponding service, which will not be repeated here. Lengthening the end-to-end service delay of the first model corresponding service, for example, the first device can adjust the terminal side model for terminal inference and the network side model for network inference according to the calculation time length of the first network side model, so that the data amount of the inference result of the adjusted terminal side model for terminal inference is greater than the data amount of the inference result of the terminal side model for terminal inference before adjustment, thereby lengthening the transmission delay of the inference result of the terminal side model for terminal inference, which is helpful to the inference accuracy of the first model. In the embodiment of the present application, the terminal side model for terminal inference and the network side model for network inference are adjusted to shorten the end-to-end service delay of the first model corresponding service.

[0123] The first device adjusts the terminal side model for terminal inference and the network side model for network inference according to the calculation time length of the first network side model to shorten the end-to-end service delay of the first model corresponding service, including the following several ways:

[0124] Way 1: The first device determines the correspondence between the bandwidth range and the model group (for example, the first association information) according to the calculation time length of the first network side model. Wherein, the first device determines the first association information according to the calculation time length of the first network side model can be to modify the correspondence between the bandwidth range and the model group (for example, the second association information) currently being used, or can also be to generate the first association information according to the calculation time length of the first network side model. Wherein, the second association information can be stored in the first device, or can also be obtained by the first device from the aforementioned cooperative inference decision point, which is not limited in the embodiment of the present application.

[0125] As described above, when the first model includes multiple sets of distributed deployment schemes, the terminal and the first node can determine the model group for inference according to the correspondence between the bandwidth range and the model group, so the terminal side model for terminal inference and the network side model for network inference can be adjusted by adjusting the correspondence between the bandwidth range and the model group.

[0126] Optionally, the first device can determine a maximum communication delay according to a calculation duration of the first network-side model and an SLA requirement of the service corresponding to the first model, and adjust the correspondence between the bandwidth range and the model group according to the maximum communication delay. The maximum communication delay is, for example, a difference between an end-to-end service delay indicated by the SLA requirement and a calculation delay of the first model. However, considering that the calculation duration of the terminal-side model is much smaller than the calculation duration of the network-side model, and the first device is difficult to obtain the calculation duration of the terminal-side model, the first device can take the calculation duration of the first network-side model as the calculation duration of the first model to determine the maximum communication delay.

[0127] In addition, since the data amount of the inference result of the terminal-side model included in different model groups is fixed, that is, the to-be-transmitted data amount corresponding to different model groups is fixed, the first device can determine the minimum bandwidth (that is, the lower bound of the bandwidth shown in Table 1) corresponding to different model groups according to the to-be-transmitted data amount and the maximum communication delay of different model groups, and determine the upper bound of the bandwidth corresponding to adjacent model groups according to the lower bound of the bandwidth corresponding to different model groups. Taking Table 1 as an example, the lower bound of the bandwidth of the model group 2 is the upper bound of the bandwidth of the model group 1, and the lower bound of the bandwidth of the model group 3 is the upper bound of the bandwidth of the model group 2. It can be understood that the model group 3 has no upper bound of the bandwidth because there is no lower bound of the bandwidth with a larger bandwidth, that is, the model group 3 is used for inference when the air interface bandwidth is greater than the lower bound of the bandwidth corresponding to the model group 3; and the model group 1 is the model group with the smallest to-be-transmitted data amount, that is, the model group 1 is used for inference regardless of the air interface bandwidth, so the lower bound of the bandwidth of the model group 1 can be the minimum value of the allowed bandwidth, for example, 0 bps or 1 bps, and it is not necessary to determine it again according to the maximum communication delay and the to-be-transmitted data amount corresponding to the model group 1. It can be understood that in the scenario of shortening the end-to-end service delay, the maximum communication delay determined according to the SLA requirement and the calculation duration of the first model is smaller than the current communication delay, so the lower bound of the bandwidth corresponding to different model groups in the first association information is greater than the lower bound of the bandwidth corresponding to the corresponding model group in the correspondence between the bandwidth range and the model group currently being used (for example, the second association information).

[0128] Taking Table 1 as an example of the second association information. The lower bounds of the bandwidth corresponding to the model group 1, the model group 2 and the model group 3 determined by the first device according to the to-be-transmitted data amount and the maximum communication delay are, for example, 1 bps, 30 bps and 60 bps respectively, and the first association information can refer to Table 2, for example.

[0129] Table 2

[0130] As can be seen from Table 1 and Table 2, if the air interface bandwidth perceived by the inference collaborative decision point is 25 bps, the inference collaborative decision point determines to use model group 2 for inference according to Table 1, and determines to use model group 1 for inference according to Table 2. That is, when the bandwidth is fixed, the first device can modify the model group used for inference by modifying the correspondence between the bandwidth range and the model group, so as to reduce the data amount of the inference result of the terminal side model, thereby shortening the communication delay and guaranteeing the end-to-end service delay of the first model corresponding service.

[0131] Optionally, the first association information can also be used to indicate the correspondence between the bandwidth range and the collaboration level, and the collaboration level is associated with the model group, for example, the collaboration level is the identification information of the model group. The first association information can refer to Table 3 for example.

[0132] Table 3

[0133] Wherein, A is the identification of model group 1, B is the identification of model group 2, and C is the identification of model group 3.

[0134] Method 2: Adjusting the model parameters of the terminal side model (i.e., the aforementioned first terminal side model) used for terminal inference and the model parameters of the network side model (i.e., the aforementioned first network side model) used for network inference according to the calculation time length of the first network side model. Wherein, the data amount of the inference result of the first terminal side model after the model parameter adjustment is less than the data amount of the inference result of the first terminal side model before the model parameter adjustment.

[0135] Wherein, the model parameters can be the codec used by the first terminal side model and the first network side model, etc. For example, adjusting the model parameters of the first terminal side model can be modifying the codec information in the sub-model file (e.g., the third sub-model file) corresponding to the first terminal side model, and the third sub-model file after the model parameter adjustment is the first sub-model file described below; adjusting the model parameters of the first network side model can be modifying the codec information in the sub-model file (e.g., the fourth sub-model file) corresponding to the first network side model, and the fourth sub-model file after the model parameter adjustment is the second sub-model file described below.

[0136] For example, the first device can determine the maximum communication delay according to the first time length and the end-to-end service delay indicated by the SLA requirement of the first model corresponding service, determine the maximum data amount of the inference result of the first terminal side model according to the maximum communication delay and the air interface bandwidth, and determine the codec used by the first terminal side model and the first network side model according to the maximum data amount.

[0137] Optionally, before adjusting the model parameters of the first terminal-side model and the first network-side model, the first device can also obtain a third sub-model file and a fourth sub-model file. For example, the first device can obtain a model file of the first model (e.g., a second model file). For example, the first device can store the second model file locally after deploying the first terminal-side model and the first network-side model, and the first device can obtain the second model file from the local storage; or the second model file can also come from a third party (e.g., a second device). When the second model file comes from the second device, the second model file can be carried in the third request mentioned above, or the first device can send a request to the second device to request the second model file, and the second model file can be carried in the response information (e.g., the third information) of the second device. After obtaining the second model file, the first device can split the second model file based on the file splitting manner corresponding to the model processing manner of the first terminal-side model and the first network-side model to obtain the third sub-model file and the fourth sub-model file.

[0138] Alternatively, the first device can also directly obtain the third sub-model file and the fourth sub-model file. For example, the first device can store the third sub-model file corresponding to the first terminal-side model and the fourth sub-model file corresponding to the first network-side model locally after deploying the first terminal-side model and the first network-side model, and the first device can obtain the third sub-model file and the fourth sub-model file from the local storage; or the third sub-model file and the fourth sub-model file can also come from a third party (e.g., a second device). When the third sub-model file and the fourth sub-model file come from the second device, the third sub-model file and the fourth sub-model file can be carried in the third request mentioned above, or the first device can send a request to the second device to request the third sub-model file and the fourth sub-model file, and the third sub-model file and the fourth sub-model file can be carried in the response information (e.g., the fourth information) of the second device.

[0139] Method 3: determining a new distributed deployment scheme according to the calculation time length of the first network-side model, deploying a new terminal-side model (e.g., a second terminal-side model) at the terminal according to the new distributed deployment scheme, and deploying a new network-side model (e.g., a second network-side model) at the first node. The data amount of the inference result of the second terminal-side model is less than the data amount of the inference result of the first terminal-side model.

[0140] The first device can redivide the first model according to the calculation time length of the first network side model to obtain the second terminal side model and the second network side model. For example, the first device can determine a maximum communication time delay according to the first time length and an end-to-end service time delay indicated by an SLA requirement of a service corresponding to the first model, determine a maximum data volume of an inference result of the first terminal side model according to the maximum communication time delay and an air interface bandwidth, determine a model processing manner according to the maximum data volume, divide the first model using the model processing manner to obtain the second terminal side model and the second network side model. The model processing manner determined according to the calculation time length of the first network side model is different from the model processing manner used to obtain the first terminal side model and the first network side model.

[0141] Optionally, the first device can deploy the second terminal side model on a terminal and the second network side model on the first node according to the new distributed deployment scheme, and further obtain a sub-model file (for example, a first sub-model file) corresponding to the second terminal side model and a sub-model file (for example, a second sub-model file) corresponding to the second network side model. For example, the first device can obtain a model file (for example, a second model file) of the first model. For example, the first device can store the second model file locally after deploying the first terminal side model and the first network side model, and the first device can obtain the second model file from the local storage. Alternatively, the second model file can come from a third party (for example, a second device). When the second model file comes from the second device, the second model file can be carried in the third request described above, or the first device can send a request to the second device for the second model file, and the second model file can be carried in the response information (for example, the third information) of the second device.

[0142] After obtaining the second model file, the first device can process (for example, split) the second model file according to the new distributed deployment scheme to obtain the first sub-model file and the second sub-model file.

[0143] The first device can send the first model file or the first association information of the first model according to the first time length, which can be related to the distributed deployment scheme included in the first model. For example, if the first model includes multiple sets of distributed deployment schemes (that is, the first device adjusts the terminal side model used for terminal inference and the network side model used for network inference according to the calculation time length of the first network side model to shorten the end-to-end service time delay of the service corresponding to the first model in the above manner 1), the first device can send the first association information according to the first time length. The first device can send the first association information according to the first time length, which can be understood as the first device can determine the first association information according to the first time length and send the first association information.

[0144] Optionally, the first device can send the first association information to the terminal and / or the first node. The first device can send the first association information to the terminal and the first node for the terminal and the first node to determine the model group for inference; or, as mentioned above, the terminal or the first node can serve as a collaborative inference decision point, and the first device can send the first association information to the collaborative inference decision point for the collaborative inference decision point to determine the model group for inference.

[0145] If the first model includes a set of distributed deployment schemes (i.e., the first device adjusts the terminal-side model for the terminal to perform inference and the network-side model for the network to perform inference according to the calculation time length of the first network-side model to shorten the end-to-end service time delay of the first model corresponding service in the above-mentioned manner 2 or manner 3), the first device can send the first model file of the first model according to the first time length. Wherein, the first device sending the first model file of the first model according to the first time length can also be understood as that the first device can determine the first sub-model file and the second sub-model file corresponding to the first model according to the first time length, and send the first sub-model file and the second sub-model file.

[0146] Optionally, the first device can send the first information including the first sub-model file to the terminal, and send the second information including the second sub-model file to the first node. Wherein, the first sub-model file may, for example, be the third sub-model file after modification of the model parameters as mentioned above, and the second sub-model file may, for example, be the fourth sub-model file after modification of the model parameters as mentioned above; or, the first sub-model file can also be a model file of a second terminal-side model for the terminal to perform inference obtained by the first device splitting the second model file according to a new distributed deployment scheme, and the second sub-model file can be a model file of a second network-side model for the network to perform inference obtained by the first device splitting the second model file according to the new distributed deployment scheme.

[0147] Optionally, if the first sub-model file is the third sub-model file after modification of the model parameters as mentioned above, and the second sub-model file is the fourth sub-model file after modification of the model parameters as mentioned above, after the first device sends the first information to the terminal and sends the second information to the first node, the first device can further send a first request to the terminal for requesting the terminal to delete the third sub-model file, and send a second request to the first node for requesting the first node to delete the fourth sub-model file.

[0148] In the above technical solution, when the end-to-end service time delay of the first model corresponding service does not meet the SLA requirement, the terminal-side model for the terminal to perform inference and the network-side model for the network to perform inference can be adjusted to reduce the amount of data to be transmitted and shorten the communication time delay of the first model corresponding service, thereby guaranteeing the end-to-end service time delay of the first model corresponding service.

[0149] Two embodiments are introduced as follows by referring to FIG. 4 and FIG. 5, which are two examples of the communication method introduced in the embodiment shown in FIG. 3.

[0150] Please refer to FIG. 4, which is a flow chart of the first example of the communication method provided by the embodiment shown in FIG. 3. In the example shown in FIG. 4, multiple sets of distributed deployment schemes are taken as an example of the first model. In the embodiment of the present application, optional steps are indicated by dashed lines.

[0151] S401: The second device sends a third request to the first device. Correspondingly, the first device receives the third request.

[0152] The third request is used to request optimization of the SLA of the service corresponding to the first model, and the third request includes the identifier of the first model, which can be the APP ID or service ID of the service corresponding to the first model.

[0153] The first device can determine the deployment node (for example, the first node mentioned above) used to deploy the network-side model corresponding to the first model according to the identifier of the first model, and perform S402.

[0154] S402: The first device sends a request for obtaining the calculation duration of the first network-side model to the first node. Correspondingly, the first node receives the request.

[0155] The calculation duration may, for example, be the average duration spent by the first network-side model in the last N times of performing an inference task, or may also be the duration spent by the first network-side model in the last time of performing an inference task, and the embodiment of the present application does not limit the calculation duration. Optionally, the request can also include the identifier of the first model, which can be an application identifier (application identifier, APP ID) or a service identifier (service ID).

[0156] As described in S301, the time or condition at which the first device obtains the calculation duration of the first network-side model can be receiving the third request of the second device, can be that the first device determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, or can be periodic acquisition, so S401 is an optional step.

[0157] S403: The first node sends the calculation duration of the first network-side model to the first device. Correspondingly, the first device receives the calculation duration of the first network-side model.

[0158] S404: The first device determines a maximum communication delay according to the calculation duration of the first network-side model and the SLA requirement of the service corresponding to the first model, and determines the first association information according to the maximum communication delay.

[0159] The related description of S404 can refer to the related description of S302 about the first device determining the first association information, which is not repeated here.

[0160] S405: The first device sends the first association information to the collaborative inference decision point. Correspondingly, the collaborative inference decision point receives the first association information.

[0161] The collaborative inference decision point may be a terminal for deploying the first terminal-side model, or may also be the first node for deploying the first network-side model.

[0162] S406: The collaborative inference decision point determines a model group for inference based on the first association information.

[0163] Please refer to FIG. 5, which is a flow chart of a second example of the communication method of the embodiment shown in FIG. 3. In the example shown in FIG. 5, it is taken as an example that the first model includes a set of distributed deployment schemes. In the embodiments of the present application, optional steps are indicated by dashed lines.

[0164] S501: The second device sends a third request to the first device. Correspondingly, the first device receives the third request.

[0165] The third request is used to request to optimize the SLA of the service corresponding to the first model, and the third request includes the identifier of the first model, which can be the APP ID or service ID of the service corresponding to the first model.

[0166] The first device can determine the deployment node (for example, the first node mentioned above) for deploying the network-side model corresponding to the first model according to the identifier of the first model, and perform S502.

[0167] S502: The first device sends a request for obtaining the calculation duration of the first network-side model to the first node. Correspondingly, the first node receives the request.

[0168] The calculation duration may be, for example, the average duration of the first network-side model for performing the inference task for the last N times, or may also be the duration of the first network-side model for performing the inference task for the last time, and the embodiments of the present application do not limit the calculation duration. Optionally, the request can also include the identifier of the first model, which can be the application identifier (application identifier, APP ID) or service identifier (service ID).

[0169] In S301, the first device obtains the time or condition of the calculation duration of the first network-side model. The time or condition of the calculation duration of the first network-side model can be receiving the third request of the second device, can be that the first device determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, or can be periodic acquisition. Therefore, S501 is an optional step.

[0170] S503: The first node sends the calculation duration of the first network-side model to the first device. Correspondingly, the first device receives the calculation duration of the first network-side model.

[0171] S504: The first device adjusts the model parameters of the first terminal-side model and the first network-side model or determines a new distributed deployment scheme according to the calculation duration of the first network-side model.

[0172] The related description of S504 can refer to the related description of S302 about the first device adjusting the model parameters of the first terminal-side model and the first network-side model or determining a new distributed deployment scheme, which will not be repeated here.

[0173] S505: The first device sends the first information to the terminal and sends the second information to the first node. Correspondingly, the terminal receives the first information, and the first node receives the second information.

[0174] The first information includes the first sub-model file, and the second information includes the second sub-model file. The related description of the first sub-model file and the second sub-model file can refer to the related description of the first sub-model file and the second sub-model file in S302, which will not be repeated here. It can be understood that the first device sending the first sub-model file to the terminal can be that the first device sends the first sub-model file to the second device, and the terminal can obtain the first sub-model file from the second device.

[0175] S506: The terminal redeploys the terminal-side model for inference according to the first sub-model file, and the first node redeploys the network-side model for inference according to the second sub-model file.

[0176] FIG. 6 is a flow diagram of a second communication method provided by an embodiment of the present application. The method includes the following steps:

[0177] S601: The second device sends a third request to the first device, or the first device determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model.

[0178] The third request is used to request to optimize the SLA of the service corresponding to the first model. The third request includes the identifier of the first model. The identifier of the first model can be the APP ID or service ID of the service corresponding to the first model.

[0179] The first device determines whether the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model in a manner, for example, that the first device determines according to the calculation duration and the communication duration of the first model, and the end-to-end service delay indicated by the SLA requirement of the service corresponding to the first model. The communication duration may, for example, be reported by a terminal or a first node.

[0180] Alternatively, the first device determines whether the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model in a manner that the first device determines according to indication information of another device. For example, if the first device receives indication information from another device indicating that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, it can be determined that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model; if the first device does not receive indication information from the other device indicating that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, it can be determined that the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model. The other device may, for example, be indicated by at least one of a terminal, a first node, or a second device, and the embodiments of the present application do not limit the manner in which the first device determines whether the SLA of the service corresponding to the first model meets the SLA requirement of the service corresponding to the first model.

[0181] The first device receives the third request or determines that the SLA of the service corresponding to the first model does not meet the SLA requirement of the service corresponding to the first model, indicating that the first node for deploying the network-side model (e.g., the first network-side model) of the first model may have more computing tasks and insufficient computing resources, and therefore S602 can be performed.

[0182] S602: The first device determines a second node with the largest remaining resources among the plurality of nodes, and deploys the first network-side model in the second node.

[0183] Optionally, the plurality of nodes can periodically report remaining resource information, and the first device can determine the second node with the largest remaining resources according to the last reported remaining resource information of the plurality of nodes; or, before performing S602, the second device can also send a request (e.g., a sixth request) to the plurality of nodes for requesting remaining resource information, and receive a response message from the plurality of nodes, the response message including the remaining resource information of the corresponding node, and the first device can determine the second node with the largest remaining resources according to the response message of the plurality of nodes.

[0184] After the first device determines the second node, the first device can send a fourth request to the second node, the fourth request being used to request the second node to deploy the first network-side model, the fourth request comprising the identifier of the first model and the model file of the first network-side model. The identifier of the first model can refer to the related description of the identifier of the first model in S301, and will not be described here.

[0185] In addition, the first device can also send a fifth request to the first node, the fifth request being used to request the first node to delete the model file of the first network-side model, the fifth request comprising the identifier of the first model.

[0186] In the above technical solution, when the first device receives the SLA used to request the optimization model to correspond to the service, or determines that the SLA of the first model corresponding to the service does not meet the SLA requirement of the service, the network-side model of the first model is deployed on the node with the largest remaining resource, which helps to reduce the calculation delay (or also referred to as inference delay) of the first network-side model, thereby guaranteeing the end-to-end service delay in the end-network collaborative inference scenario.

[0187] FIG. 7 shows a structural schematic diagram of a communication device provided by an embodiment of the present application. The communication device 700 can be a first device or circuit system of the first device according to the embodiments of any one of FIGS. 3-6, used to implement the method corresponding to the first device in the above method embodiments. Alternatively, the communication device 700 can be a terminal or circuit system of the terminal according to the embodiments of any one of FIGS. 3-6, used to implement the method corresponding to the terminal in the above method embodiments. Alternatively, the communication device 700 can be a first node or circuit system of the first node according to the embodiments of any one of FIGS. 3-6, used to implement the method corresponding to the first node in the above method embodiments. Alternatively, the communication device 700 can be a second device or circuit system of the second device according to the embodiments of any one of FIGS. 3-6, used to implement the method corresponding to the second device in the above method embodiments. For example, one circuit system is a chip system.

[0188] The communication device 700 comprises at least one processor 701. The processor 701 can be used for internal processing of the device, realizing certain control processing functions. Optionally, the processor 701 comprises instructions. Optionally, the processor 701 can store data. Optionally, different processors can be independent devices, can be located in different physical positions, and can be located on different integrated circuits. Optionally, different processors can be integrated in one or more processors, for example, integrated on one or more integrated circuits.

[0189] Optionally, the communication device 700 comprises one or more memories 703 to store instructions. Optionally, the memories 703 can also store data. The processor and the memories can be separately arranged or integrated together.

[0190] Optionally, the communication device 700 comprises a communication line 702 and at least one communication interface 704. Since the memories 703, the communication line 702 and the communication interface 704 are optional, they are all represented by dashed lines in FIG. 7.

[0191] Optionally, the communication device 700 can also comprise a transceiver and / or an antenna. The transceiver can be used to send information to other devices or receive information from other devices. The transceiver can be referred to as a transceiver, a transceiving circuit, an input / output interface, etc., and is used to realize the transceiving function of the communication device 700 through the antenna. Optionally, the transceiver comprises a transmitter and a receiver. Illustratively, the transmitter can be used to generate a radio frequency signal from a baseband signal, and the receiver can be used to convert a radio frequency signal into a baseband signal.

[0192] The processor 701 can comprise a general central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of programs of the solutions of the present application.

[0193] The communication line 702 can comprise a path to transmit information between the above-mentioned components.

[0194] The communication interface 704 uses any transceiver-like device for communicating with other devices or communication networks, such as an Ethernet, a radio access network (RAN), a wireless local area networks (WLAN), a wired access network, etc.

[0195] The memory 703 can be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM), or other type of dynamic storage device that can store information and instructions for execution by the processor 701, and can be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, a magnetic disk storage or other magnetic storage devices, or any other medium capable of storing instructions or data that can be accessed by a computer, but is not limited to this. The memory 703 can exist independently, and be connected to the processor 701 through the communication line 702. Alternatively, the memory 703 can be integrated with the processor 701.

[0196] The memory 703 is configured to store computer-executable instructions for implementing the solutions of the present application, and the processor 701 is configured to control the execution of the computer-executable instructions stored in the memory 703. The processor 701 is configured to execute the computer-executable instructions stored in the memory 703, so as to implement the steps performed by the first device, the terminal, the first node, or the second device in the embodiments shown in any one of FIGS. 3-6.

[0197] Optionally, the computer-executable instructions in the embodiments of the present application can also be referred to as application program codes, and the embodiments of the present application are not limited in this regard.

[0198] In a specific implementation, as an example, the processor 701 can include one or more CPUs, such as CPU0 and CPU1 in FIG. 7.

[0199] In a specific implementation, as an example, the communication device 700 can include multiple processors, such as the processor 701 and the processor 705 in FIG. 7. Each of these processors can be a single-CPU processor or a multi-CPU processor. The processor herein can refer to one or more devices, circuits, and / or processing cores for processing data (for example, computer program instructions).

[0200] When the apparatus shown in Fig. 7 is a chip, for example, a chip of the first apparatus, the terminal, the first node or the second apparatus, the chip comprises the processor 701 (and can comprise the processor 705), the communication line 702 and the communication interface 704, and optionally, the chip can comprise the memory 703. Specifically, the communication interface 704 can be an input interface, a pin or a circuit, etc. The memory 703 can be a register, a cache, etc. The processor 701 and the processor 705 can be a general CPU, a microprocessor, an ASIC, or one or more integrated circuits for executing programs for controlling the communication method of any of the above embodiments.

[0201] The embodiments of the present application can divide the functional modules of the apparatus according to the above method examples. 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 module can be realized in the form of hardware or in the form of a software functional module. It should be noted that the division of the modules in the embodiments of the present application is illustrative, and is only a logical function division. When each functional module is divided according to each function, for example, Fig. 8 shows an apparatus schematic diagram. The apparatus 800 can be the first apparatus, the terminal, the first node or the second apparatus, or a chip of the first apparatus, the terminal, the first node or the second apparatus, involved in each of the above method embodiments. The apparatus 800 comprises a sending unit 801, a processing unit 802 and a receiving unit 803.

[0202] It should be understood that the apparatus 800 can be used to implement the steps performed by the first apparatus, the terminal, the first node or the second apparatus in the communication method of the embodiments of the present application. The related features can refer to any of the embodiments shown in Figs. 3-6, and will not be described here.

[0203] Optionally, the functions / implementation processes of the sending unit 801, the receiving unit 803 and the processing unit 802 in Fig. 8 can be realized by the processor 701 in Fig. 7 invoking computer-executable instructions stored in the memory 703. Alternatively, the functions / implementation processes of the processing unit 802 in Fig. 8 can be realized by the processor 701 in Fig. 7 invoking computer-executable instructions stored in the memory 703, and the functions / implementation processes of the sending unit 801 and the receiving unit 803 in Fig. 8 can be realized by the communication interface 704 in Fig. 7.

[0204] Optionally, when the apparatus 800 is a chip or a circuit, the functions / implementation processes of the sending unit 801 and the receiving unit 803 can also be realized by a pin or a circuit, etc.

[0205] The application further provides a computer readable storage medium, which stores a computer program or instructions, and when the computer program or instructions are executed, the method executed by the first device, the terminal, the first node or the second device in the foregoing method embodiments is implemented. In this way, the functions described in the foregoing embodiments can be implemented in the form of software function units and sold or used as independent products. Based on this understanding, the technical solutions of the application can be embodied in the form of a software product in essence or in the part that contributes to the application or part of the technical solutions. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, a first device, a terminal, a first node or a second device, etc.) to execute all or part of the steps of the methods described in the embodiments of the application. The storage medium includes a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk, and various media that can store program codes.

[0206] The application further provides a computer program product, which includes computer program codes, and when the computer program codes are executed on a computer, the computer is caused to execute the method executed by the first device, the terminal, the first node or the second device in any of the foregoing method embodiments.

[0207] The embodiments of the application further provide a processing device, which includes a processor and an interface; the processor is used to execute the method executed by the first device, the terminal, the first node or the second device related to any of the foregoing method embodiments.

[0208] In the embodiments described above, all or some of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or some of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded into and executed by a computer, all or some of the procedures or functions described in the embodiments of the present application are performed. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from a website, a computer, a server or a data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media sets. The available media can be a magnetic medium (for example, floppy disk, hard disk, magnetic tape), an optical medium (for example, DVD), or a semiconductor medium (for example, solid state disk (SSD)) and the like.

[0209] The various illustrative logical blocks, modules, circuits, and algorithms described in connection with the embodiments disclosed herein can be implemented or performed by a general purpose processor, a digital signal processor (DSP), an ASIC, a field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the described functions. The general purpose processor can be a microprocessor, optionally, the general purpose processor can also be any conventional processor, controller, microcontroller, or state machine. The processor can also be implemented by a combination of computing devices, such as a digital signal processor and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other similar configuration.

[0210] The steps of methods or algorithms described in the embodiments of the present application can be directly embedded in hardware, a software unit executed by a processor, or a combination of both. The software unit can be stored in a RAM, a flash memory, a ROM, an erasable programmable read-only memory (EPROM), an EEPROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium in the art. The storage medium can be connected to the processor, so that the processor can read information from, and write information to, the storage medium. Alternatively, the storage medium can be integrated into the processor. The processor and the storage medium can be located in an ASIC, which can be located in the first device, the terminal, the first node, or the second device. Alternatively, the processor and the storage medium can also be located in different components of the first device, the terminal, the first node, or the second device.

[0211] These computer program instructions can also be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce computer-implemented processing, so that the instructions executed by the computer or other programmable device provide steps for implementing the functions specified in one or more flows in the flowcharts and / or one or more blocks in the block diagrams.

[0212] The contents in various embodiments of the present application can be mutually referred to, and the terms and / or descriptions in different embodiments are consistent and can be mutually referred to if there is no special description and logical conflict. The technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0213] It can be understood that the first device, the terminal, the first node, or the second device in the embodiments of the present application can perform part or all of the steps in the embodiments of the present application, and these steps or operations are only examples, and other operations or various modifications of the operations can also be performed in the embodiments of the present application. In addition, each step can be performed in a different order from the order presented in the embodiments of the present application, and it is possible that not all the operations in the embodiments of the present application are performed.

Claims

1. A communication method characterized by comprising: The method applied to a first device comprises: obtaining a calculation duration of a first network side model, the first network side model being a sub-model of a first model processed for network inference, the first network side model corresponding to a first terminal side model, the first terminal side model being a sub-model of the first model processed for terminal inference; sending a first model file or first association information of the first model according to the calculation duration; wherein the first association information is used to indicate a corresponding relationship between a bandwidth range and a model group, one model group including one network side model and one terminal side model, the one network side model and the one terminal side model corresponding.

2. The method of claim 1, wherein, sending the first model file or the first association information of the first model according to the calculation duration comprises: sending the first association information to a terminal and / or a first node according to the calculation duration, the first node being a node deploying the first network side model.

3. The method of claim 2, wherein, sending the first association information to the terminal and / or the first node according to the calculation duration comprises: determining a maximum communication delay according to the calculation duration and service level agreement (SLA) requirements of a service corresponding to the first model; determining the first association information according to the maximum communication delay; sending the first association information to the terminal and / or the first node.

4. The method according to any one of claims 1 to 3, characterized in that, the first association information is used to indicate a corresponding relationship between a bandwidth range and a model group, comprising: the first association information is used to indicate a corresponding relationship between a bandwidth range and a cooperation level, the cooperation level being associated with a model group.

5. The method of claim 1, wherein, the first model file includes a first sub-model file corresponding to a second terminal side model and a second sub-model file corresponding to a second network side model, the second terminal side model corresponding to the second network side model, sending the first model file or the first association information of the first model according to the calculation duration comprises: sending first information including the first sub-model file to a terminal and sending second information including the second sub-model file to a first node according to the calculation duration, the first node being a node deploying the first network side model.

6. The method of claim 5, wherein, the method further comprises: obtaining a third sub-model file corresponding to the first terminal side model and a fourth sub-model file corresponding to the first network side model; adjusting model parameters in the third sub-model file to obtain the first sub-model file according to the calculation duration, and adjusting model parameters in the fourth sub-model file to obtain the second sub-model file according to the calculation duration.

7. The method of claim 6, wherein, adjusting model parameters in the third sub-model file according to the calculation duration, and adjusting model parameters in the fourth sub-model file according to the calculation duration, comprises: determining an SLA of a service corresponding to the first model according to the calculation duration; adjusting the model parameters in the third sub-model file according to the SLA and SLA requirements of the service, and adjusting the model parameters in the fourth sub-model file according to the SLA and SLA requirements of the service.

8. The method of claim 6 or 7, wherein, The method further includes: The method further includes:

9. The method of claim 8, wherein, The method further includes: The method further includes:

10. The method of claim 6 or 7, wherein, The method further includes: The method further includes:

11. The method of claim 5, wherein, The method further includes: The method further includes: The method further includes:

12. The method of claim 11, wherein, The method further includes: The method further includes:

13. The method according to any one of claims 5 to 12, characterized in that, The method further includes: The method further includes: The method further includes:

14. The method according to any one of claims 1 to 13, characterized in that, The method further includes: The method further includes:

15. A method of communication, comprising: The method further includes: The method further includes: The method further includes:

16. The method of claim 15, wherein, The method further includes: The method further includes:

17. The method of claim 15 or 16, wherein, The method further includes: The method further includes:

18. A communications device, characterized by The method further includes:

19. A communications device, characterized by The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further includes: The method further 20. A computer-readable storage medium, characterized in that, The computer-readable storage medium is for storing a computer program which, when executed on a computer, causes the method according to any one of claims 1 to 14 to be performed, or causes the method according to any one of claims 15 to 17 to be performed.

21. A computer program product, characterised in that, The computer program product comprises a computer program which, when executed on a computer, causes the computer to perform the method according to any one of claims 1 to 14, or causes the computer to perform the method according to any one of claims 15 to 17.

Images