Collaborative reasoning method and communication device

Abstract

The present application provides a collaborative reasoning method and communication device, which relate to the field of communication technology and can reduce the latency of a terminal device in obtaining a target reasoning result. The method includes: the terminal device determines a first reasoning result based on a first machine learning (ML) sub-model. The first ML sub-model is a part of the ML model. Thereafter, the terminal device sends the first reasoning result. Then, the terminal device receives a target reasoning result. The target reasoning result is the reasoning result of the ML model determined based on the first reasoning result.

Description

Technical Field

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

[0002] A machine learning (ML) model is a mathematical model or signal model composed of training data and expert knowledge, used to statistically describe the characteristics of a given data set. When ML models are introduced into wireless communication networks, the specific implementation methods are as follows:

[0003] When a terminal device stores an ML model, it determines the inference result based on its own data and the stored ML model, and then performs relevant processing based on the inference result. For example, in a remote driving scenario, the terminal device acts as an onboard module, onboard module, onboard component, onboard chip, or onboard unit built into the vehicle. The terminal device adjusts the vehicle's driving conditions based on the inference result.

[0004] However, terminal devices lack high computing power and cannot meet the latency requirements of real-world applications. For example, the latency of remote driving services cannot exceed 5ms. When the ML model is implemented as an AlexNet model, the terminal device's computing power must be at least 39G floating point operations per second (FLOPS). However, the terminal device's computing power cannot meet these requirements, which increases the latency of the terminal device in obtaining inference results.

[0005] In summary, when ML reasoning is introduced into wireless communication networks, the problem of "long delay in obtaining inference results" cannot be solved for terminal devices. Summary of the Invention

[0006] The embodiments of the present application provide a collaborative reasoning method and communication device that can reduce the latency for a terminal device to obtain a target reasoning result and improve the security of the terminal device data.

[0007] To achieve the above objectives, the present invention adopts the following technical solutions:

[0008] In a first aspect, an embodiment of the present application provides a collaborative reasoning method, wherein the execution subject of the method can be a terminal device or a chip used in the terminal device. The following description takes the execution subject being a terminal device as an example. The method includes: the terminal device determines a first reasoning result based on a first machine learning ML sub-model. The first ML sub-model is part of the ML model. Thereafter, the terminal device sends the first reasoning result, and then the terminal device receives a target reasoning result. The target reasoning result is the reasoning result of the ML model determined based on the first reasoning result.

[0009] In this way, the terminal device uses the first ML sub-model to perform partial inference operations to obtain a first inference result. After the terminal device sends the first inference result, the first network device calculates all the information of the first inference result in combination with the target ML sub-model to obtain the target inference result, and then provides it to the terminal device. This eliminates the need for the terminal device to perform a complete inference operation, reducing the latency for the terminal device to obtain the target inference result. In addition, the terminal device provides the first network device with the intermediate results of the ML model calculation rather than the input data of the ML model, reducing the risk of "data privacy exposure" and improving the security of the terminal device data.

[0010] In one possible design, when the terminal device is connected to the first network device before determining the first inference result, the terminal device sends the first inference result, including: the terminal device sends all information about the first inference result to the first network device. The terminal device receives the target inference result, including: the terminal device receives the target inference result from the first network device, where the target inference result is an inference result of the ML model determined based on all information about the first inference result. In other words, the terminal device is connected to the first network device before performing local inference, provides the first inference result to the first network device, and then obtains the inference result from the first network device.

[0011] In one possible design, the terminal device obtains information of the first ML sub-model, including: the terminal device receives information of the first ML sub-model from the first network device, so that the terminal device performs local reasoning.

[0012] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives the first model information from the first network device, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position, the first candidate indication information and the first segmentation position are at least one, a first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with a first candidate indication information. The terminal device determines the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information. That is, the first network device sends the first target indication information (such as the segmentation option corresponding to the first ML sub-model to indicate the segmentation position of the ML model) to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0013] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the terminal device sends reasoning requirement information to the first network device, wherein the reasoning requirement information includes an identifier of the ML model and time information when the terminal device obtains the target reasoning result, and the reasoning requirement information is used to determine information of the first ML sub-model. Here, because the reasoning requirement information includes the time information when the terminal device obtains the target reasoning result, after the first ML sub-model is determined based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0014] In one possible design, before a terminal device sends a first inference result, it accesses a first network device, and while the terminal device is sending the first inference result, it accesses a second network device. The terminal device sends the first inference result, including: the terminal device sends a first portion of information about the first inference result to the first network device, where the first network device is the network device that the terminal device accessed before accessing the second network device; and the terminal device sends a second portion of information about the first inference result to the second network device. The terminal device receives a target inference result, including: the terminal device receives a target inference result from the second network device, where the target inference result is an inference result of an ML model determined based on the first portion of information and the second portion of information. That is, after the terminal device sends the first portion of information about the first inference result to the first network device, the terminal device accesses the second network device (e.g., the terminal device switches from the first network device to the second network device), and the terminal device no longer interacts with the first network device, thereby sending the second portion of information about the first inference result to the second network device. Furthermore, the terminal device obtains the target inference result from the second network device.

[0015] In one possible design, when the terminal device accesses the first network device before sending the first inference result and accesses the second network device during the process of the terminal device sending the first inference result, the terminal device sends the first inference result, including: the terminal device sends all the information of the first inference result to the first network device, wherein the first network device is the network device accessed before the terminal device accesses the second network device. The terminal device receives the target inference result, including: the terminal device receives the target inference result from the second network device, wherein the target inference result is the inference result of the ML model determined based on all the information of the first inference result. That is, after the terminal device sends the complete first inference result to the first network device, the terminal device accesses the second network device (such as the terminal device switches, that is, switches from the first network device to the second network device) to obtain the target inference result from the second network device.

[0016] In one possible design, when the terminal device connects to the second network device before sending the first inference result, the terminal device sends the first inference result, including: the terminal device sends all information about the first inference result to the second network device. The terminal device receives the target inference result, including: the terminal device receives the target inference result from the second network device, where the target inference result is an inference result of the ML model determined based on all information about the first inference result. In other words, after the terminal device obtains the first inference result, the terminal device has connected to the second network device, provides the first inference result to the second network device, and then obtains the inference result from the second network device.

[0017] In one possible design, when the terminal device accesses the first network device before determining the first reasoning result, the collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives information of the first ML sub-model from the first network device, so that the terminal device performs local reasoning.

[0018] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives the first model information from the first network device, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position, and the first candidate indication information and the first segmentation position are at least one. A first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with a first candidate indication information. The terminal device determines the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information. That is, the first network device sends the first target indication information (i.e., the segmentation option corresponding to the first ML sub-model to indicate the segmentation position of the ML model) to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0019] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the terminal device sending reasoning requirement information to the first network device, wherein the reasoning requirement information includes time information when the terminal device obtains the target reasoning result, and the reasoning requirement information is used to determine information of the first ML sub-model. Here, because the reasoning requirement information includes the time information when the terminal device obtains the target reasoning result, after determining the first ML sub-model based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0020] In one possible design, when the terminal device accesses the second network device before determining the first inference result, the terminal device sends the first inference result, including: the terminal device sends all information of the first inference result to the second network device. The terminal device receives a target inference result, including: the terminal device receives the target inference result from the second network device, where the target inference result is an inference result of the ML model determined based on all information of the first inference result.

[0021] In one possible design, when the terminal device accesses the second network device from the first network device through switching, the collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives information of the first ML sub-model from the first network device, so that the terminal device performs local reasoning.

[0022] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of the embodiment of the present application also includes: the terminal device receives the first model information from the first network device, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position, and the first candidate indication information and the first segmentation position are at least one. A first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with a first candidate indication information. The terminal device determines the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information. That is, the first network device sends the first target indication information to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0023] In one possible design, when the terminal device accesses the second network device through the radio resource control RRC connection re-establishment or RRC connection recovery process, the collaborative reasoning method of the embodiment of the present application also includes: the terminal device receives information of the first ML sub-model from the second network device, so that the terminal device performs local reasoning.

[0024] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of the embodiment of the present application also includes: the terminal device receives the first model information from the second network device, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position, and the first candidate indication information and the first segmentation position are at least one. A first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with a first candidate indication information. The terminal device determines the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information. That is, the second network device sends the first target indication information to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0025] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the terminal device sending reasoning requirement information to the first network device, wherein the reasoning requirement information includes time information when the terminal device obtains the target reasoning result, and the reasoning requirement information is used to determine information of the first ML sub-model. Here, because the reasoning requirement information includes the time information when the terminal device obtains the target reasoning result, after determining the first ML sub-model based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0026] In one possible design, the input data for the first ML sub-model is data generated by the terminal device. Here, the terminal device obtains the inference results of the first ML sub-model based on its own generated data and then provides the network device with the intermediate results of the ML model calculation rather than the input data of the ML model. This reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0027] In the second aspect, an embodiment of the present application provides a collaborative reasoning method, the execution subject of the method can be a first network device, or it can be a chip applied to the first network device. The following description is taken as an example in which the execution subject is the first network device. The method includes: the first network device receives first reasoning information from the terminal device. The first reasoning information includes all or part of the information of the first reasoning result, the first reasoning result is the reasoning result of the first machine learning ML sub-model, and the first ML sub-model is a part of the ML model. Afterwards, the first network device sends the second reasoning information to the second network device. The second reasoning information is determined based on the first reasoning information, and the second reasoning information is used to determine the target reasoning result of the ML model, or the second reasoning information is the target reasoning result.

[0028] In this way, after the first network device receives the first reasoning information from the terminal device, it sends the second reasoning information to the second network device, so that the second network device determines the target reasoning result and then provides it to the terminal device. Alternatively, the second reasoning information is the target reasoning result, which is transmitted to the second network device. Here, the first reasoning information is determined based on the first reasoning result. The first reasoning result is the reasoning result obtained by the terminal device using the first ML sub-model to perform a partial reasoning operation, so that the terminal device does not need to perform a complete reasoning operation, thereby reducing the delay for the terminal device to obtain the target reasoning result. In addition, the terminal device provides the first network device with the intermediate result calculated by the ML model instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0029] In one possible design, the collaborative reasoning method of the embodiment of the present application further includes: the first network device determines information about the first ML sub-model. Thereafter, the first network device sends the information about the first ML sub-model to the terminal device, so that the terminal device performs an inference operation.

[0030] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: a first network device receiving reasoning requirement information from a terminal device. The reasoning requirement information includes time information when the terminal device obtains a target reasoning result. The first network device determining information about the first ML sub-model includes: the first network device determining information about the first ML sub-model based on the reasoning requirement information.

[0031] That is, the information of the first ML sub-model is determined based on the inference requirement information to meet the latency requirement of the terminal device to obtain the target inference result. In the case where the first ML sub-model is determined by the first network device, the terminal device provides the inference requirement information to the first network device.

[0032] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of the embodiment of the present application also includes: the first network device sends the first model information to the terminal device. The first model information includes the correspondence between the first candidate indication information and the first segmentation position. There is at least one first candidate indication information and one first segmentation position, and one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with one first candidate indication information. The first model information and the first target indication information are used by the terminal device to determine the first ML sub-model, which saves transmission resources compared to transmitting the full information of the first ML sub-model.

[0033] In one possible design, the first reasoning information includes all the information of the first reasoning result. The collaborative reasoning method of the embodiment of the present application also includes: the first network device determines the target reasoning result based on all the information of the first reasoning result and the target ML sub-model. Among them, the second reasoning information is the target reasoning result, and the input data of the target ML sub-model corresponds to the output data of the first ML sub-model. That is, the first network device performs an inference operation based on the first reasoning result to obtain the target reasoning result, and transmits it to the second network device, thereby reducing the amount of computing of the terminal device and the second network device.

[0034] In one possible design, the first reasoning information includes all the information of the first reasoning result. The collaborative reasoning method of the embodiment of the present application also includes: the first network device determines the second reasoning result based on the all the information of the first reasoning result and the second ML sub-model. The second reasoning information is the second reasoning result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model. That is to say, the first network device performs a partial reasoning operation based on the first reasoning result to obtain the second reasoning result, and transmits it to the second network device, so that the second network device continues to perform the reasoning operation based on the second reasoning result, thereby reducing the amount of calculation of the terminal device.

[0035] In one possible design, the collaborative reasoning method of the embodiment of the present application further includes: the first network device sending information of a target ML sub-model to the second network device. The input data of the target ML sub-model corresponds to the output data of the second ML sub-model, and the target ML sub-model is used by the second network device to determine a target reasoning result.

[0036] Here, when the first network device performs local reasoning and obtains the second reasoning result but does not obtain the target reasoning result, the first network device also provides the target ML sub-model to the second network device, so that the second network device performs reasoning based on the target ML sub-model and obtains the target reasoning result.

[0037] In one possible design, the first reasoning information is identical to the second reasoning information. The collaborative reasoning method of the embodiment of the present application further includes: the first network device sending information about a target ML sub-model to the second network device. The input data of the target ML sub-model corresponds to the output data of the first ML sub-model, and the target ML sub-model is used by the second network device to determine a target reasoning result.

[0038] Here, when the first network device forwards the first inference information to the second network device, the first network device also provides information of the target ML sub-model to the second network device, so that the second network device performs inference based on the target ML sub-model to obtain a target inference result.

[0039] In one possible design, the information of the target ML sub-model includes second target indication information. The collaborative reasoning method of the embodiment of the present application also includes: the first network device receives the second model information from the second network device. The second model information includes the correspondence between the second candidate indication information and the second segmentation position, and there is at least one second candidate indication information and second segmentation position. One second candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the second segmentation position that has a corresponding relationship with one second candidate indication information. The first network device determines the second target indication information from the second candidate indication information based on the correspondence between the second candidate indication information and the second segmentation position, and the target ML sub-model, which saves transmission resources compared to transmitting the full information of the target ML sub-model.

[0040] In a third aspect, an embodiment of the present application provides a collaborative reasoning method, the execution subject of which can be a second network device or a chip applied to the second network device. The following description is made using the second network device as an example. The method includes: the second network device obtains third reasoning information. The third reasoning information is determined based on all information of the first reasoning result, and the first reasoning result is an inference result after the operation of the first machine learning ML sub-model, and the first ML sub-model is a part of the ML model. Afterwards, the second network device sends the target reasoning result to the terminal device. The target reasoning result is the reasoning result of the ML model determined based on the third reasoning information.

[0041] In this way, since the third reasoning information is determined based on all the information of the first reasoning result, and the first reasoning result is the reasoning result obtained by the terminal device using the first ML sub-model to perform a partial reasoning operation. After the second network device obtains the third reasoning information, the second network device can send the target reasoning result to the terminal device. Here, the target reasoning result is determined based on the third reasoning information, so that the terminal device does not need to perform a complete reasoning operation, which reduces the delay for the terminal device to obtain the target reasoning result. In addition, the terminal device provides the first network device with the intermediate result of the ML model calculation instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0042] In one possible design, when the terminal device is connected to the second network device before the second network device obtains the third reasoning information, the third reasoning information is all the information of the first reasoning result. The second network device obtains the third reasoning information, including: the second network device receives all the information of the first reasoning result from the terminal device. The collaborative reasoning method of the embodiment of the present application also includes: the second network device determines the target reasoning result based on all the information of the first reasoning result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model. That is to say, when the terminal device is connected to the second network device, the second network device obtains all the information of the first reasoning result from the terminal device to perform network-side operations and obtain the target reasoning result, thereby reducing the amount of operations of the terminal device.

[0043] In one possible design, the second network device sends information of the first ML sub-model, including: the second network device sends the information of the first ML sub-model to the terminal device, so that the terminal device performs an inference operation.

[0044] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: a second network device receives reasoning requirement information from a terminal device, wherein the reasoning requirement information includes time information for the terminal device to obtain a target reasoning result. The second network device determines information about the first ML sub-model based on the reasoning requirement information. Here, since the reasoning requirement information includes time information for the terminal device to obtain the target reasoning result, after the first ML sub-model is determined based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0045] In one possible design, when the terminal device is connected to the second network device during the process of the second network device acquiring the third reasoning information, the third reasoning information is all the information of the first reasoning result. The second network device acquires the third reasoning information, including: the second network device receives the first part of the information of the first reasoning result from the terminal device; the second network device receives the second part of the information of the first reasoning result from the first network device. The collaborative reasoning method of the embodiment of the present application also includes: the second network device determines the target reasoning result based on the first part of the information, the second part of the information and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0046] That is, after the terminal device sends the first portion of the first inference result to the first network device, the terminal device connects to the second network device and no longer interacts with the first network device, thereby sending the second portion of the first inference result to the second network device. Furthermore, the second network device can also obtain the first portion of the first inference result from the first network device to perform network-side inference and obtain the target inference result.

[0047] In one possible design, when the terminal device accesses the second network device after the second network device obtains the third inference information, the third inference information is all the information of the first inference result. The second network device obtains the third inference information, including: the second network device receives all the information of the first inference result from the first network device. The collaborative inference method of the embodiment of the present application also includes: the second network device determines the target inference result based on all the information of the first inference result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0048] That is to say, after the terminal device sends the complete first reasoning result to the first network device, the terminal device connects to the second network device. In this case, the second network device obtains all information of the first reasoning result from the first network device to perform local reasoning and obtain the target reasoning result.

[0049] In one possible design, when the terminal device is connected to the second network device by the first network device before the second network device obtains the third inference information, the third inference information is all the information of the first inference result. The second network device obtains the third inference information, including: the second network device receives all the information of the first inference result from the terminal device. The collaborative inference method of the embodiment of the present application also includes: the second network device determines the target inference result based on the all the information of the first inference result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0050] That is, after the terminal device obtains the first reasoning result, the terminal device has been connected to the second network device, and the terminal device provides the first reasoning result to the second network device so that the second network device performs network-side reasoning to obtain the target reasoning result.

[0051] In one possible design, the third reasoning information is the second reasoning result, the second reasoning result is the reasoning result of the second ML sub-model determined based on all the information of the first reasoning result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model. The second network device obtains the third reasoning information, including: the second network device receives the second reasoning result from the first network device. The collaborative reasoning method of the embodiment of the present application also includes: the second network device determines the target reasoning result based on the second reasoning result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the second ML sub-model.

[0052] That is, when the first network device performs an inference operation and obtains a second inference result, the second network device obtains the second inference result from the first network device, and continues to perform an inference operation based on the second inference result to obtain a target inference result.

[0053] In one possible design, after the second network device obtains the information of the target ML sub-model, when the terminal device is connected to the second network device, the second network device obtains the information of the target ML sub-model, including: the second network device receives the information of the target ML sub-model from the first network device to perform inference operations and obtain the target inference results.

[0054] In one possible design, the information of the target ML sub-model includes second target indication information. The collaborative reasoning method of the embodiment of the present application further includes: the second network device sends second model information to the first network device, wherein the second model information includes a correspondence between second candidate indication information and a second split position; there is at least one second candidate indication information and at least one second split position, one second candidate indication information indicates that the ML model is to be split, and the position where the ML model is to be split is a second split position that has a corresponding relationship with one second candidate indication information; the second model information is used by the first network device to determine the second target indication information.

[0055] That is, when the first network device indicates the target ML sub-model to the second network device through the second target indication information, the second network device provides the second model information to the first network device, so that the first network device determines the second target indication information from the second model information, thereby saving transmission resources.

[0056] In one possible design, the third reasoning information is a target reasoning result. The second network device obtains the third reasoning information, including: the second network device receives the target reasoning result from the first network device.

[0057] That is, when the first network device performs an inference operation and obtains an inference result, the second network device obtains the target inference result from the first network device.

[0058] In one possible design, the second network device sends the information of the first ML sub-model, including: the second network device sends the information of the first ML sub-model to the terminal device; or, the second network device sends the information of the first ML sub-model to the first network device.

[0059] Here, when the terminal device accesses the second network device through the RRC connection recovery process or the RRC connection re-establishment process, the second network device sends the information of the first ML sub-model to the terminal device, so that the terminal device performs the inference operation. When the terminal device accesses the second network device through the handover process, the second network device sends the information of the first ML sub-model to the first network device, so that the first network device provides the information of the first ML sub-model to the terminal device, thereby enabling the terminal device to perform the inference operation.

[0060] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the second network device receives reasoning requirement information from the first network device, wherein the reasoning requirement information includes time information for the terminal device to obtain the target reasoning result. The second network device determines the information of the first ML sub-model based on the reasoning requirement information. Here, the second network device obtains the reasoning requirement information from the first network device. Since the reasoning requirement information includes the time information for the terminal device to obtain the target reasoning result, after determining the first ML sub-model based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0061] In a fourth aspect, an embodiment of the present application provides a collaborative reasoning method, the execution subject of which can be a terminal device or a chip applied to the terminal device. The following description is made using the terminal device as an example. When the access network device is implemented as a separate architecture, the method includes: the terminal device determines a first reasoning result based on a first machine learning ML sub-model. The first ML sub-model is a part of the ML model. Afterwards, the terminal device sends the first reasoning result, and then the terminal device receives the target reasoning result. The target reasoning result is the reasoning result of the ML model determined based on the first reasoning result.

[0062] In this way, the terminal device uses the first ML sub-model to perform partial inference operations to obtain a first inference result. After the terminal device sends the first inference result, the first distribution unit DU combines the target ML sub-model to calculate all the information of the first inference result to obtain the target inference result, and then provides it to the terminal device, so that the terminal device does not need to perform a complete inference operation, reducing the latency of the terminal device obtaining the target inference result. In addition, the terminal device provides the first DU with the intermediate results of the ML model calculation instead of the input data of the ML model, reducing the risk of "data privacy exposure" and improving the security of the terminal device data.

[0063] In one possible design, when the terminal device accesses the first DU before determining the first inference result, the terminal device sends the first inference result, including: the terminal device sends all information of the first inference result to the first DU. The terminal device receives the target inference result, including: the terminal device receives the target inference result from the first DU, wherein the target inference result is an inference result of the ML model determined based on all information of the first inference result. That is, the terminal device has accessed the first DU before performing local inference, and the terminal device provides the first inference result to the first DU, and then obtains the inference result from the first DU.

[0064] In one possible design, the terminal device obtains information of the first ML sub-model, including: the terminal device receives information of the first ML sub-model from the first DU, so that the terminal device performs local inference.

[0065] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives the first model information from the first DU, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position, the first candidate indication information and the first segmentation position are at least one, a first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with a first candidate indication information. The terminal device determines the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information. That is, the first DU sends the first target indication information (i.e., the segmentation option corresponding to the first ML sub-model to indicate the segmentation position of the ML model) to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0066] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the terminal device sends reasoning requirement information to the first DU, wherein the reasoning requirement information includes the identifier of the ML model and the time information when the terminal device obtains the target reasoning result, and the reasoning requirement information is used to determine the information of the first ML sub-model. Here, since the reasoning requirement information includes the time information when the terminal device obtains the target reasoning result, after the first ML sub-model is determined based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement of the terminal device to obtain the target reasoning result.

[0067] In one possible design, before the terminal device sends the first inference result, the first DU is accessed, and when the terminal device is in the process of accessing the second DU, the terminal device sends the first inference result, including: the terminal device sends the first part of the information of the first inference result to the first DU, wherein the first DU is the DU accessed before the terminal device accesses the second DU; the terminal device sends the second part of the information of the first inference result to the second DU. The terminal device receives the target inference result, including: the terminal device receives the target inference result from the second DU, wherein the target inference result is an inference result of the ML model determined based on the first part of the information and the second part of the information. That is, after the terminal device sends the first part of the information of the first inference result to the first DU, the terminal device accesses the second DU (such as the terminal device switches, that is, switches from the first DU to the second DU), and the terminal device no longer interacts with the first DU, thereby sending the second part of the information of the first inference result to the second DU. In addition, the terminal device obtains the target inference result from the second DU.

[0068] In one possible design, when the terminal device accesses the first DU before sending the first reasoning result, and accesses the second DU after the terminal device sends the first reasoning result and before receiving the target reasoning result, the terminal device sends the first reasoning result, including: the terminal device sends all the information of the first reasoning result to the first DU, wherein the first DU is the DU accessed before the terminal device accesses the second DU. The terminal device receives the target reasoning result, including: the terminal device receives the target reasoning result from the second DU, wherein the target reasoning result is an inference result of the ML model determined based on all the information of the first reasoning result. That is, after the terminal device sends the complete first reasoning result to the first DU, the terminal device accesses the second DU (such as the terminal device switches, that is, switches from the first DU to the second DU) to obtain the target reasoning result from the second DU.

[0069] In one possible design, when the terminal device accesses the second DU before sending the first inference result, the terminal device sends the first inference result, including: the terminal device sends all information of the first inference result to the second DU. The terminal device receives the target inference result, including: the terminal device receives the target inference result from the second DU, wherein the target inference result is an inference result of the ML model determined based on all information of the first inference result. That is, after the terminal device obtains the first inference result, the terminal device has accessed the second DU, the terminal device provides the first inference result to the second DU, and then obtains the inference result from the second DU.

[0070] In one possible design, when the terminal device accesses the first DU before determining the first reasoning result, the collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives information of the first ML sub-model from the first DU, so that the terminal device performs local reasoning.

[0071] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives the first model information from the first DU, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position, and the first candidate indication information and the first segmentation position are at least one. A first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with a first candidate indication information. The terminal device determines the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information. That is, the first DU sends the first target indication information (i.e., the segmentation option corresponding to the first ML sub-model to indicate the segmentation position of the ML model) to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0072] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the terminal device sends reasoning requirement information to the first DU, wherein the reasoning requirement information includes time information for the terminal device to obtain the target reasoning result, and the reasoning requirement information is used to determine information of the first ML sub-model. Here, since the reasoning requirement information includes the time information for the terminal device to obtain the target reasoning result, after the first ML sub-model is determined based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0073] In one possible design, when the terminal device accesses the second DU before determining the first inference result, the terminal device sends the first inference result, including: the terminal device sends all information of the first inference result to the second DU. The terminal device receives a target inference result, including: the terminal device receives the target inference result from the second DU, where the target inference result is an inference result of the ML model determined based on all information of the first inference result.

[0074] In one possible design, the collaborative reasoning method of the embodiment of the present application further includes: the terminal device receiving information of the first ML sub-model from the first DU. Here, when the terminal device switches from the first DU to the second DU, the terminal device obtains the information of the first ML sub-model through the first DU.

[0075] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of an embodiment of the present application also includes: the terminal device receives the first model information from the first DU, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position, and the first candidate indication information and the first segmentation position are at least one. A first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with a first candidate indication information. The terminal device determines the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information. That is, the first DU sends the first target indication information (i.e., the segmentation option corresponding to the first ML sub-model to indicate the segmentation position of the ML model) to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0076] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the terminal device sends reasoning requirement information to the first DU, wherein the reasoning requirement information includes time information for the terminal device to obtain the target reasoning result, and the reasoning requirement information is used to determine information of the first ML sub-model. Here, since the reasoning requirement information includes the time information for the terminal device to obtain the target reasoning result, after the first ML sub-model is determined based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0077] In one possible design, the input data for the first ML sub-model is data generated by the terminal device. Here, the terminal device obtains the inference results of the first ML sub-model based on its own generated data and then provides the DU with the intermediate results of the ML model calculation rather than the input data of the ML model. This reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0078] In a fifth aspect, an embodiment of the present application provides a collaborative reasoning method, the execution subject of the method can be a first DU, or it can be a chip applied to the first DU. The following description is taken as an example in which the execution subject is the first DU. The method includes: the first DU receives first reasoning information from a terminal device. The first reasoning information includes all or part of the information of the first reasoning result, the first reasoning result is the reasoning result of the first machine learning ML sub-model, and the first ML sub-model is a part of the ML model. Afterwards, the first DU sends the second reasoning information to the second DU. The second reasoning information is determined based on the first reasoning information, and the second reasoning information is used to determine the target reasoning result of the ML model, or the second reasoning information is the target reasoning result.

[0079] In this way, after the first DU receives the first reasoning information of the terminal device, the second reasoning information is sent to the second DU, so that the second DU determines the target reasoning result and then provides it to the terminal device. Alternatively, the second reasoning information is the target reasoning result, which is transmitted to the second DU. Here, the first reasoning information is determined based on the first reasoning result. The first reasoning result is the reasoning result obtained by the terminal device using the first ML sub-model to perform partial reasoning operations, so that the terminal device does not need to perform a complete reasoning operation, reducing the delay for the terminal device to obtain the target reasoning result. In addition, the terminal device provides the first DU with the intermediate results calculated by the ML model instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0080] In one possible design, the collaborative reasoning method of the embodiment of the present application further includes: the first DU determining information of the first ML sub-model. Thereafter, the first DU sends the information of the first ML sub-model to the terminal device so that the terminal device performs the reasoning operation.

[0081] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the first DU receiving reasoning requirement information from the terminal device. The reasoning requirement information includes time information when the terminal device obtains the target reasoning result. When the first DU determines the first ML sub-model, the first DU determines the first ML sub-model based on the reasoning requirement information.

[0082] That is, the first ML sub-model is determined based on the inference requirement information to meet the latency requirement of the terminal device for obtaining the target inference result. In the case where the first ML sub-model is determined by the first DU, the terminal device provides the inference requirement information to the first DU.

[0083] In one possible design, the information of the first ML sub-model includes first target indication information. The collaborative reasoning method of the embodiment of the present application also includes: the first network device sends the first model information to the terminal device. The first model information includes the correspondence between the first candidate indication information and the first segmentation position. There is at least one first candidate indication information and one first segmentation position, and one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with one first candidate indication information. The first model information and the first target indication information are used by the terminal device to determine the first ML sub-model, which saves transmission resources compared to transmitting the full information of the first ML sub-model.

[0084] In one possible design, the first reasoning information includes all the information of the first reasoning result. The collaborative reasoning method of the embodiment of the present application also includes: the first DU determines the target reasoning result based on all the information of the first reasoning result and the target ML sub-model. Among them, the second reasoning information is the target reasoning result, and the input data of the target ML sub-model corresponds to the output data of the first ML sub-model. That is to say, the first DU performs an inference operation based on the first reasoning result to obtain the target reasoning result, and transmits it to the second DU, thereby reducing the amount of computation of the terminal device and the second DU.

[0085] In one possible design, the first reasoning information includes all the information of the first reasoning result. The collaborative reasoning method of the embodiment of the present application also includes: the first DU determines the second reasoning result based on all the information of the first reasoning result and the second ML sub-model. The second reasoning information is the second reasoning result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model. That is to say, the first DU performs a partial reasoning operation based on the first reasoning result to obtain the second reasoning result, and transmits it to the second DU, so that the second DU continues to perform the reasoning operation based on the second reasoning result, thereby reducing the amount of calculation of the terminal device.

[0086] In one possible design, the collaborative reasoning method of the embodiment of the present application further includes: the first DU sending information of a target ML sub-model to the second DU. The input data of the target ML sub-model corresponds to the output data of the second ML sub-model, and the target ML sub-model is used by the second DU to determine the target reasoning result.

[0087] Here, when the first DU performs local reasoning to obtain the second reasoning result but fails to obtain the target reasoning result, the first DU also provides the information of the target ML sub-model to the second DU so that the second DU performs reasoning based on the target ML sub-model and obtains the target reasoning result.

[0088] In one possible design, the first reasoning information is the same as the second reasoning information. The collaborative reasoning method of the embodiment of the present application further includes: the first DU sends information about the target ML sub-model to the second DU. The input data of the target ML sub-model corresponds to the output data of the first ML sub-model, and the target ML sub-model is used by the second DU to determine the target reasoning result.

[0089] Here, in the case where the first DU forwards the first inference information to the second DU, the first DU also provides the information of the target ML sub-model to the second DU, so that the second DU performs inference based on the target ML sub-model to obtain the target inference result.

[0090] In one possible design, the information of the target ML sub-model includes second target indication information. The collaborative reasoning method of the embodiment of the present application also includes: the first DU receives the second model information from the second DU. The second model information includes the correspondence between the second candidate indication information and the second segmentation position, and there is at least one second candidate indication information and second segmentation position. One second candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the second segmentation position that has a corresponding relationship with one second candidate indication information. The first DU determines the second target indication information from the second candidate indication information based on the correspondence between the second candidate indication information and the second segmentation position, as well as the target ML sub-model, which saves transmission resources compared to transmitting the full information of the target ML sub-model.

[0091] In a sixth aspect, an embodiment of the present application provides a collaborative reasoning method, the execution subject of the method can be a second DU, or a chip applied to the second DU. The following description is taken as an example that the execution subject is the second DU. The method includes: the second DU obtains third reasoning information. The third reasoning information is determined based on all information of the first reasoning result, and the first reasoning result is the reasoning result after the terminal device calculates the first machine learning ML sub-model, and the first ML sub-model is part of the ML model. Afterwards, the second DU sends the target reasoning result to the terminal device. The target reasoning result is the reasoning result of the ML model determined based on the third reasoning information.

[0092] In this way, since the third reasoning information is determined based on all the information of the first reasoning result, and the first reasoning result is the reasoning result obtained by the terminal device using the first ML sub-model to perform partial reasoning operations. After the second DU obtains the third reasoning information, the second DU can send the target reasoning result to the terminal device. Here, the target reasoning result is determined based on the third reasoning information, so that the terminal device does not need to perform a complete reasoning operation, which reduces the latency for the terminal device to obtain the target reasoning result. In addition, the terminal device provides the first DU with the intermediate results calculated by the ML model instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0093] In one possible design, when the terminal device is connected to the second DU before the second DU obtains the third reasoning information, the third reasoning information is all the information of the first reasoning result. The second DU obtains the third reasoning information, including: the second DU receives all the information of the first reasoning result from the terminal device. The collaborative reasoning method of the embodiment of the present application also includes: the second DU determines the target reasoning result based on all the information of the first reasoning result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model. That is to say, when the terminal device is connected to the second DU, the second DU obtains all the information of the first reasoning result from the terminal device to perform operations and obtain the target reasoning result, thereby reducing the amount of operations of the terminal device.

[0094] In one possible design, the second DU sends information of the first ML sub-model, including: the second DU sends the information of the first ML sub-model to the terminal device, so that the terminal device performs an inference operation.

[0095] In one possible design, the collaborative reasoning method of the embodiment of the present application further includes: the second DU receiving reasoning requirement information from the terminal device, wherein the reasoning requirement information includes time information when the terminal device obtains the target reasoning result. The second DU determines information about the first ML sub-model based on the reasoning requirement information.

[0096] In one possible design, when the terminal device is connected to the second DU during the process of the second DU acquiring the third reasoning information, the third reasoning information is all the information of the first reasoning result. The second DU acquires the third reasoning information, including: the second DU receives the first part of the information of the first reasoning result from the terminal device; the second DU receives the second part of the information of the first reasoning result from the first DU. The collaborative reasoning method of the embodiment of the present application also includes: the second DU determines the target reasoning result based on the first part of the information, the second part of the information and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model. Here, since the reasoning requirement information includes the time information of the terminal device acquiring the target reasoning result, after the first ML sub-model is determined based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement of the terminal device to obtain the target reasoning result.

[0097] In one possible design, when the terminal device accesses the second DU after the second DU obtains the third reasoning information, the third reasoning information is all the information of the first reasoning result. The second DU obtains the third reasoning information, including: the second DU receives all the information of the first reasoning result from the first DU. The collaborative reasoning method of the embodiment of the present application also includes: the second DU determines the target reasoning result based on all the information of the first reasoning result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0098] That is, after the terminal device sends the first part of the first inference result to the first DU, the terminal device accesses the second DU and no longer interacts with the first DU, thereby sending the second part of the first inference result to the second DU. Furthermore, the second DU can also obtain the first part of the first inference result from the first DU to perform network-side inference and obtain the target inference result.

[0099] In one possible design, when the terminal device is connected to the second DU via the first DU before the second DU obtains the third inference information, the third inference information is all the information of the first inference result. The second DU obtains the third inference information, including: the second DU receives all the information of the first inference result from the terminal device. The collaborative inference method of the embodiment of the present application also includes: the second DU determines the target inference result based on all the information of the first inference result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0100] That is to say, after the terminal device obtains the first reasoning result, the terminal device has accessed the second DU, and the terminal device provides the first reasoning result to the second DU so that the second DU performs network-side reasoning to obtain the target reasoning result.

[0101] In one possible design, the third reasoning information is the second reasoning result, and the second reasoning result is the reasoning result of the second ML sub-model determined based on all the information of the first reasoning result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model. The second DU obtains the third reasoning information, including: the second DU receives the second reasoning result from the first DU. The collaborative reasoning method of the embodiment of the present application also includes: the second DU determines the target reasoning result based on the second reasoning result and the target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the second ML sub-model.

[0102] That is, when the first DU performs an inference operation and obtains a second inference result, the second DU obtains the second inference result from the first DU and continues to perform an inference operation based on the second inference result to obtain a target inference result.

[0103] In one possible design, after the second DU obtains the information of the target ML sub-model, when the terminal device accesses the second DU, the second DU obtains the information of the target ML sub-model, including: the second DU receives the information of the target ML sub-model from the first DU to perform inference operations and obtain the target inference results.

[0104] In one possible design, the information of the target ML sub-model includes second target indication information. The collaborative reasoning method of the embodiment of the present application further includes: the second DU sends second model information to the first DU, wherein the second model information includes a correspondence between second candidate indication information and a second segmentation position; there is at least one second candidate indication information and a second segmentation position, one second candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is a second segmentation position that has a corresponding relationship with one second candidate indication information; the second model information is used by the first DU to determine the second target indication information.

[0105] That is, when the first DU indicates the target ML sub-model for the second DU through the second target indication information, the second DU provides the second model information to the first DU, so that the first DU determines the second target indication information from the second model information, thereby saving transmission resources.

[0106] In one possible design, the third reasoning information is a target reasoning result. The second DU obtains the third reasoning information, including: the second DU receives the target reasoning result from the first DU.

[0107] That is, when the first DU performs an inference operation and obtains an inference result, the second DU obtains the target inference result from the first DU.

[0108] In one possible design, the second DU sends the information of the first ML sub-model, including: the second DU sends the information of the first ML sub-model to the first DU.

[0109] Here, when the terminal device accesses the second DU through the switching process, the second DU sends the information of the first ML sub-model to the first DU, so that the first DU provides the information of the first ML sub-model to the terminal device, thereby enabling the terminal device to perform inference operations.

[0110] In one possible design, the collaborative reasoning method of an embodiment of the present application further includes: the second DU receives reasoning requirement information from the first DU, wherein the reasoning requirement information includes time information for the terminal device to obtain the target reasoning result. The second DU determines the information of the first ML sub-model based on the reasoning requirement information. Here, the second DU obtains the reasoning requirement information from the first DU. Since the reasoning requirement information includes the time information for the terminal device to obtain the target reasoning result, after the first ML sub-model is determined based on the reasoning requirement information, the terminal device performs local reasoning based on the first ML sub-model, which can meet the latency requirement for the terminal device to obtain the target reasoning result.

[0111] In a seventh aspect, an embodiment of the present application provides a communication device, comprising: a unit for executing each step in the first aspect or any possible design of the first aspect, or the fourth aspect or any possible design of the fourth aspect. The communication device can be a terminal device in the first aspect or any possible design of the first aspect, or a chip that implements the functions of the terminal device; or, the communication device can be a terminal device in the fourth aspect or any possible design of the fourth aspect, or a chip that implements the functions of the terminal device. The communication device includes a module, unit, or means corresponding to the above method, and the module, unit, or means can be implemented by hardware, software, or by hardware executing the corresponding software implementation. The hardware or software includes one or more modules or units corresponding to the above functions.

[0112] In an eighth aspect, an embodiment of the present application provides a communication device, comprising: a processor and a memory; the memory is used to store computer instructions, and when the processor executes the instructions, the communication device executes the method described in the first aspect or any possible design of the first aspect, or the fourth aspect or any possible design of the fourth aspect. The communication device can be a terminal device in the first aspect or any possible design of the first aspect, or a chip that implements the functions of the terminal device; or the communication device can be a terminal device in the fourth aspect or any possible design of the fourth aspect, or a chip that implements the functions of the terminal device.

[0113] In a ninth aspect, an embodiment of the present application provides a communication device, comprising: a processor; the processor is configured to couple with a memory, and after reading instructions from the memory, execute, according to the instructions, the method described in the first aspect or any possible design of the first aspect, or the fourth aspect or any possible design of the fourth aspect. The communication device may be a terminal device in the first aspect or any possible design of the first aspect, or a chip that implements the functions of the terminal device; or the communication device may be a terminal device in the fourth aspect or any possible design of the fourth aspect, or a chip that implements the functions of the terminal device.

[0114] In a tenth aspect, an embodiment of the present application provides a chip comprising a logic circuit and an input / output interface. The input / output interface is used to communicate with a module outside the chip, for example, the input / output interface outputs first reasoning information, or the input / output interface inputs a target reasoning result. The logic circuit is used to run a computer program or instruction to implement the collaborative reasoning method provided by the first aspect or any possible design of the first aspect, or the fourth aspect or any possible design of the fourth aspect. The chip may be a chip that implements the terminal device function in the first aspect or any possible design of the first aspect; or, the chip may be a chip that implements the terminal device function in the fourth aspect or any possible design of the fourth aspect.

[0115] In an eleventh aspect, an embodiment of the present application provides a communication device, comprising: a unit for executing each step in the second aspect or any possible design of the second aspect. The communication device can be the first network device in the second aspect or any possible design of the second aspect, or a chip that implements the function of the first network device. The communication device includes a module, unit, or means corresponding to the above method, which can be implemented by hardware, software, or by executing the corresponding software implementation through hardware. The hardware or software includes one or more modules or units corresponding to the above functions.

[0116] In a twelfth aspect, an embodiment of the present application provides a communication device, comprising: a processor and a memory; the memory being configured to store computer instructions, and when the processor executes the instructions, causing the communication device to perform the method described in the second aspect or any possible design of the second aspect. The communication device may be the first network device described in the second aspect or any possible design of the second aspect, or a chip that implements the functions of the first network device.

[0117] In a thirteenth aspect, an embodiment of the present application provides a communication device, comprising: a processor; the processor being configured to be coupled to a memory, and after reading instructions from the memory, execute, according to the instructions, the method described in the second aspect or any possible design of the second aspect. The communication device may be the first network device described in the second aspect or any possible design of the second aspect, or a chip implementing the functions of the first network device.

[0118] In a fourteenth aspect, an embodiment of the present application provides a chip comprising a logic circuit and an input / output interface. The input / output interface is used to communicate with a module outside the chip. For example, the input / output interface inputs first reasoning information, or the input / output interface outputs second reasoning information. The logic circuit is used to run a computer program or instruction to implement the collaborative reasoning method provided in the second aspect or any possible design of the second aspect above. The chip can be a chip that implements the first network device function in the second aspect or any possible design of the second aspect above.

[0119] In a fifteenth aspect, an embodiment of the present application provides a communication device, comprising: a unit for executing each step in the third aspect or any possible design of the third aspect. The communication device can be the second network device in the third aspect or any possible design of the third aspect, or a chip that implements the function of the second network device. The communication device includes a module, unit, or means corresponding to the above method, which can be implemented by hardware, software, or by executing the corresponding software implementation through hardware. The hardware or software includes one or more modules or units corresponding to the above functions.

[0120] In a sixteenth aspect, embodiments of the present application provide a communication device, comprising: a processor and a memory; the memory being configured to store computer instructions, such that when the processor executes the instructions, the communication device performs the method described in the third aspect or any possible design of the third aspect. The communication device may be the second network device described in the third aspect or any possible design of the third aspect, or a chip implementing the functionality of the second network device.

[0121] In a seventeenth aspect, an embodiment of the present application provides a communication device, comprising: a processor; the processor being configured to be coupled to a memory, and after reading instructions from the memory, execute, in accordance with the instructions, the method described in the third aspect or any possible design of the third aspect. The communication device may be the second network device described in the third aspect or any possible design of the third aspect, or a chip implementing the functions of the second network device.

[0122] In an eighteenth aspect, an embodiment of the present application provides a chip comprising a logic circuit and an input / output interface. The input / output interface is used to communicate with a module outside the chip; for example, the input / output interface outputs a target reasoning result. The logic circuit is used to run a computer program or instruction to implement the collaborative reasoning method provided in the third aspect or any possible design of the third aspect. The chip may be a chip that implements the second network device function in the third aspect or any possible design of the third aspect.

[0123] In the nineteenth aspect, an embodiment of the present application provides a communication device, which includes: a unit for executing each step in the fifth aspect or any possible design of the fifth aspect. The communication device can be the first DU in the fifth aspect or any possible design of the fifth aspect, or a chip that implements the function of the first DU. The communication device includes a module, unit, or means corresponding to the above method, which can be implemented by hardware, software, or by hardware executing the corresponding software. The hardware or software includes one or more modules or units corresponding to the above functions.

[0124] In a twentieth aspect, an embodiment of the present application provides a communication device, comprising: a processor and a memory; the memory being configured to store computer instructions, and when the processor executes the instructions, causing the communication device to perform the method described in the fifth aspect or any possible design of the fifth aspect. The communication device may be the first DU described in the fifth aspect or any possible design of the fifth aspect, or a chip implementing the functionality of the first DU.

[0125] In a twenty-first aspect, an embodiment of the present application provides a communication device, comprising: a processor; the processor being configured to be coupled to a memory, and after reading instructions from the memory, execute, according to the instructions, the method described in the fifth aspect or any possible design of the fifth aspect. The communication device may be the first DU described in the fifth aspect or any possible design of the fifth aspect, or a chip implementing the functions of the first DU.

[0126] In aspect 22, an embodiment of the present application provides a chip comprising a logic circuit and an input / output interface. The input / output interface is used to communicate with a module outside the chip, for example, the input / output interface inputs first reasoning information, or the input / output interface outputs second reasoning information. The logic circuit is used to run a computer program or instruction to implement the collaborative reasoning method provided in the fifth aspect or any possible design of the fifth aspect above. The chip can be a chip that implements the first DU function in the fifth aspect or any possible design of the fifth aspect above.

[0127] In aspect 23, an embodiment of the present application provides a communication device, comprising: a unit for executing each step in the sixth aspect or any possible design of the sixth aspect. The communication device can be the second DU in the sixth aspect or any possible design of the sixth aspect, or a chip that implements the function of the second DU. The communication device includes a module, unit, or means corresponding to the above method, which can be implemented by hardware, software, or by hardware executing the corresponding software. The hardware or software includes one or more modules or units corresponding to the above functions.

[0128] In a twenty-fourth aspect, an embodiment of the present application provides a communication device, comprising: a processor and a memory; the memory being configured to store computer instructions, wherein when the processor executes the instructions, the communication device performs the method described in the sixth aspect or any possible design of the sixth aspect. The communication device may be the second DU described in the sixth aspect or any possible design of the sixth aspect, or a chip implementing the functionality of the second DU.

[0129] In a twenty-fifth aspect, an embodiment of the present application provides a communication device, comprising: a processor; the processor being configured to be coupled to a memory and, after reading instructions from the memory, execute, according to the instructions, the method described in the sixth aspect or any possible design of the sixth aspect. The communication device may be the second DU described in the sixth aspect or any possible design of the sixth aspect, or a chip implementing the functionality of the second DU.

[0130] In aspect 26, an embodiment of the present application provides a chip comprising a logic circuit and an input / output interface. The input / output interface is used to communicate with a module outside the chip, for example, the input / output interface outputs a target reasoning result. The logic circuit is used to run a computer program or instruction to implement the collaborative reasoning method provided in aspect 6 or any possible design of aspect 6 above. The chip may be a chip that implements the second DU function in aspect 6 or any possible design of aspect 6 above.

[0131] In aspect 27, an embodiment of the present application provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the collaborative reasoning method of any one of the above aspects.

[0132] In aspect 28, an embodiment of the present application provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the collaborative reasoning method of any one of the above aspects.

[0133] In aspect 29, an embodiment of the present application provides a circuit system, the circuit system including a processing circuit, and the processing circuit is configured to execute a collaborative reasoning method as described in any one of the above aspects.

[0134] In aspect 30, an embodiment of the present application provides a collaborative reasoning system, which includes a first network device and a second network device.

[0135] Among them, the technical effects brought about by any design in aspects 7 to 30 can refer to the beneficial effects of the corresponding methods provided above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0136] Figure 1 A schematic diagram of a neural network provided in an embodiment of the present application;

[0137] Figure 2 A schematic diagram of a network architecture provided in an embodiment of the present application;

[0138] Figure 3 A schematic diagram of a distributed network architecture provided in an embodiment of the present application;

[0139] Figure 4 A flowchart of the first collaborative reasoning method provided in an embodiment of the present application;

[0140] Figure 5 A schematic diagram of a process for configuring a first computing radio bearer provided in an embodiment of the present application;

[0141] Figure 6 A schematic diagram of a process for transmitting a first machine learning sub-model provided in an embodiment of the present application;

[0142] Figure 7a A layered schematic diagram of a communication protocol provided in an embodiment of the present application;

[0143] Figure 7b A layered diagram of another communication protocol provided in an embodiment of the present application;

[0144] Figure 8 A flowchart of the second collaborative reasoning method provided in an embodiment of the present application;

[0145] Figure 9a A schematic diagram of a process for configuring a target calculation radio bearer according to an embodiment of the present application;

[0146] Figure 9b A schematic diagram of a flow chart of a transmission target machine learning sub-model provided in an embodiment of the present application;

[0147] Figure 9cA layered schematic diagram of another communication protocol provided in an embodiment of the present application;

[0148] Figure 9d A layered schematic diagram of another communication protocol provided in an embodiment of the present application;

[0149] Figure 10 A schematic diagram of a process for configuring a target computing radio bearer according to another embodiment of the present application;

[0150] Figure 11 A flowchart of the third collaborative reasoning method provided in an embodiment of the present application;

[0151] Figure 12 A flowchart of the fourth collaborative reasoning method provided in an embodiment of the present application;

[0152] Figure 13 A schematic diagram of a process for transmitting a first machine learning sub-model according to another embodiment of the present application;

[0153] Figure 14 A schematic diagram of a process for configuring a target computing radio bearer according to an embodiment of the present application;

[0154] Figure 15 A layered schematic diagram of another communication protocol provided in an embodiment of the present application;

[0155] Figure 16 A schematic diagram of the structure of a communication device provided in an embodiment of the present application;

[0156] Figure 17 A structural diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0157] The terms "first" and "second" in the description and drawings of this application are used to distinguish different objects or to distinguish different processing of the same object, rather than to describe a specific order of objects. In addition, the terms "including" and "having" and any variations thereof mentioned in the description of this application are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products, or devices. In the embodiments of this application, "multiple" includes two or more. In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner. In the embodiments of this application, "transmit" includes "send" or "receive."

[0158] First, let’s introduce the technical terms involved in this application:

[0159] 1. Switch

[0160] In a wireless communication system, when a terminal device moves from one cell to another, or due to network reasons, business load adjustment, equipment failure, etc., the terminal device may switch from the source cell to the target cell to ensure the continuity of communication between the terminal device and the network. The above process is called "switching". Here, before the switching, the access network device that communicates with the terminal device is described as the source access network device. After the switching, the access network device that communicates with the terminal device is described as the target access network device. In an embodiment of the present application, the source access network device is described as the "first network device" and the target access network device is described as the "second network device".

[0161] 2. Radio resource control (RRC) inactive state and RRC connected state

[0162] Both the RRC inactive state and the RRC connected state are used to describe the status of a terminal device.

[0163] For a terminal device in the RRC inactive state, the user plane bearer of the air interface is suspended, while the user plane bearer and control plane bearer between the access network device and the core network device are still maintained. The terminal device stores the access layer context and supports cell reselection. When the terminal device initiates a call or service request, the user plane bearer of the air interface needs to be activated, and the existing user plane bearer and control plane bearer between the access network device and the core network device need to be reused.

[0164] For terminal devices in RRC connected state, the control plane bearer of the air interface has been established.

[0165] Here, an access network device that converts a terminal device from an RRC connected state to an RRC inactive state, or an access network device that stores the access layer context of the terminal device, is described as a source access network device. The access network device to which the terminal device in the RRC inactive state reselects or the access network device to which the terminal device newly accesses is described as a target access network device. In an embodiment of the present application, the source access network device is described as a "first network device" and the target access network device is described as a "second network device."

[0166] 3. RRC connection recovery (resume)

[0167] When the terminal device is in the RRC inactive state and needs to update the radio access network notification area (RNA), the terminal device sends an RRC resume request message to the second network device. Accordingly, the second network device receives the RRC resume request message from the terminal device. Afterwards, the second network device sends information such as the radio bearer configuration to the terminal device, enabling the terminal device to perform data transmission. The above process is called "RRC connection recovery."

[0168] 4. RRC connection reestablishment

[0169] The purpose of RRC connection reestablishment is to enable a terminal device in the RRC connected state to restore the RRC connection when an RRC connection anomaly occurs, thereby reducing the impact of the anomaly on communications. The terminal device initiates RRC connection reestablishment when at least one of the following situations occurs: first, radio link failure; second, integrity check failure; third, RRC connection reconfiguration failure.

[0170] 5. ML Model

[0171] ML model, also known as artificial intelligence (AI) model. ML model is a mathematical model or signal model composed of training data and expert knowledge, used to statistically describe the characteristics of a given data set. ML models include supervised learning models, unsupervised learning models, reinforcement learning models, neural network models, etc. For example, Figure 1 A typical neural network model is shown. The neural network model includes multiple neurons, such as Figure 1 As shown in the circle in . The neural network model includes 1 input layer (such as Figure 1 ), 3 hidden layers (as shown in the circle filled with diagonal lines), Figure 1 ) and 1 output layer (shown as a blank circle in Figure 1 Indicated by the circle filled with vertical lines). Among them, the input layer receives external input signals, the hidden layer and the output layer process the input signals at different stages, and the final result is output by the output layer. Each layer of the neural network model includes at least one neuron. Each neuron receives input signals from other neurons, and these input signals are transmitted through weighted connections. The neuron first compares the total input value received with the threshold of the neuron, and then generates the output of the neuron through the activation function. In addition, by increasing the data of the hidden layer in the ML model, and / or increasing the number of neurons in the hidden layer, the accuracy of the ML model can be improved, or the capacity of the ML model can be increased. Here, only the neural network model is taken as an example to illustrate the structure of the ML model. Supervised learning models, unsupervised learning models or reinforcement learning models, etc. are similar to Figure 1 The neural network models shown have the same structure: an input layer, a hidden layer, and an output layer. The connections between adjacent layers differ for supervised learning models, unsupervised learning models, or reinforcement learning models. Furthermore, hidden layers can also be described as "intermediate layers."

[0172] Generally, ML is divided into training and inference. Training is the process of learning an ML model to perform a specific task based on a training dataset. Inference is the process of computing the ML model on the input data to produce inference results.

[0173] When an ML model is introduced into a wireless communication network, the related art shows the following two possible implementation methods:

[0174] Method 1: The terminal device stores the ML model. The terminal device determines the inference result based on its own data and the stored ML model.

[0175] Method 2: The network device stores the ML model. The terminal device sends input data to the network device. The network device determines the inference result based on the input data provided by the terminal device and its stored ML model. The network device sends the inference result to the terminal device, allowing the terminal device to obtain the inference result.

[0176] However, in the first approach, the terminal device must have high computing power to meet the latency requirements of actual services. In the second approach, the terminal device does not need to perform ML reasoning, so the computing power requirements of the terminal device are relatively low. However, the terminal device provides input data to the network device, and this input data belongs to the terminal device, which may expose the data privacy of the terminal device.

[0177] In summary, when ML reasoning is introduced into wireless communication networks, the problem of long latency in obtaining inference results cannot be solved for terminal devices. Furthermore, some related technologies still cannot address the issue of data privacy exposure.

[0178] In view of this, an embodiment of the present application provides a collaborative reasoning method, which is applicable to various communication systems. The collaborative reasoning method provided in the embodiment of the present application can be applied to a long-term evolution (LTE) system, a fifth-generation (5G) communication network, or other similar networks, or other future networks. Figure 2 The following is a schematic diagram of the architecture of a communication system applicable to the collaborative reasoning method according to an embodiment of the present application. The communication system may include an access network device 21, a terminal device 20 communicating with the access network device 21, and a core network device 22 communicating with the access network device 21. The number of the terminal device 20, the access network device 21, and the core network device 22 may be one or more. Figure 2 Only one terminal device 20, two access network devices 21 and one core network device 22 are shown. Figure 2 It is only a schematic diagram and does not constitute a limitation on the applicable scenarios of the collaborative reasoning method in the embodiment of the present application.

[0179] The terminal device 20, also known as user equipment (UE), mobile station (MS), mobile terminal (MT), etc., is a device that provides voice / data connectivity to users, such as a handheld device or vehicle-mounted device with wireless connection capabilities. The terminal device can specifically be: a mobile phone, a tablet computer, a laptop computer, a PDA, a mobile internet device (MID), a wearable device, a virtual reality (VR) device, an augmented reality (AR) device, a wireless terminal in industrial control, a wireless terminal in self-driving, a wireless terminal in remote medical surgery, a wireless terminal in a smart grid, a wireless terminal in transportation safety, a wireless terminal in a smart city, a wireless terminal in a smart home, a terminal device in a 5G communication network or a communication network after 5G, etc., and the embodiments of the present application are not limited thereto.

[0180] Among them, the core network device 22 is a device deployed in the core network to provide services for the terminal device 20. In systems using different wireless access technologies, the names of core network devices with similar wireless communication functions may be different. For example, when the collaborative reasoning method of the embodiment of the present application is applied to the 5G system, the core network device can be, for example, but not limited to, the access and mobility management function (AMF) or the network data analytics function (NWDAF). Among them, the AMF has the functions of mobility management, registration management, connection management, lawful interception, support for transmission of session management (SM) information between the terminal device 20 and the session management function (SMF), access authentication and access authorization. NWDAF can collect data from various network functions (NF), application functions (AF), operation administration and maintenance (OAM) to perform network function analysis and prediction. For the convenience of description only, in the embodiment of the present application, the above-mentioned devices that can provide services for the terminal device 20 are collectively referred to as core network devices. The interface between the core network device and the access network device is the NG interface.

[0181] The access network device 21 is a device in a wireless communication network, such as a radio access network (RAN) node that connects the terminal device 20 to the wireless communication network. Currently, some examples of RAN nodes include: next generation Node B (gNB), next generation evolutionary Node B (ng-eNB) connected to the next generation core network, transmission reception point (TRP), evolved Node B (eNB), radio network controller (RNC), Node B (NB), base station controller (BSC), base transceiver station (BTS), home base station (e.g., home evolved Node B, or HNB), base band unit (BBU), or wireless fidelity (Wi-Fi) access point (AP).

[0182] In one possible embodiment, the access network device 21 in the embodiment of the present application includes a centralized unit (CU) and a distributed unit (DU), such as Figure 3As shown. Among them, the number of CUs and DUs can be one or more. CU and DU can be understood as the division of the access network device 21 from the perspective of logical functions. Among them, CU and DU can be physically separated or deployed together, and the embodiment of the present application does not make specific restrictions on this. The CU and DU can be connected through an interface, for example, it can be an F1 interface. CU and DU can be divided according to the protocol layer of the wireless network. For example, the functions of the radio resource control (RRC) layer, the service data adaptation protocol (SDAP) layer, and the packet data convergence protocol (PDCP) layer are set in the CU, while the functions of the radio link control (RLC) layer, the media access control (MAC) layer, the physical (PHY) layer, etc. are set in the DU. It can be understood that the division of the CU and DU processing functions according to this protocol layer is only an example, and can also be divided in other ways. The embodiment of the present application does not make specific restrictions on this.

[0183] Optionally, the CU includes a CU control plane (CU control plane, CU-CP) and a CU user plane (CU user plane, CU-UP). Among them, a CU includes a CU-CP and one or more CU-UPs. CU-CP and CU-UP can be understood as a division of the CU from the perspective of logical functions. CU-CP and CU-UP can be divided according to the protocol layer of the wireless network. For example, the control plane in the RRC layer and the PDCP layer is set in the CU-CP, and the user plane in the PDCP layer is set in the CU-UP. In addition, the function of the SDAP layer may also be set in the CU-UP. The CU-CP and CU-UP can be connected through an interface, such as an E1 interface. The CU-CP and DU can be connected through the control plane interface (F1-C) of F1, and the CU-UP and DU can be connected through the user plane interface (F1-U) of F1. Furthermore, the CU, DU or CU-CP can be connected to the data analysis and management (DAM) unit through the G1 interface respectively. Optionally, the DAM unit can be used as an internal function of the CU, DU or CU-CP. In this case, the G1 interface is an internal interface.

[0184] Understandably, the above Figure 2The communication system shown is only for the purpose of more clearly illustrating the technical solution of the embodiment of the present application, and does not constitute a limitation on the technical solution provided by the embodiment of the present application. For example, the communication system may also include other devices, such as a network control device ( Figure 2 (Not shown). The network control device may be an operation administration and maintenance (OAM) system, which may also be called a network management system. The network control device may manage the access network devices and core network devices.

[0185] The communication system and service scenarios described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that, with the evolution of network architecture and the emergence of new service scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0186] The collaborative reasoning method provided in the embodiments of the present application is described in detail below.

[0187] It should be noted that the message names between the network elements or the names of the parameters in the messages in the following embodiments of the present application are only examples, and other names may be used in specific implementations. They are uniformly described here and will not be repeated below.

[0188] In an embodiment of the present application, the terminal device provides reasoning-related information (such as a first reasoning result) to the first network device, and receives a target reasoning result from the first network device. Here, on the terminal device side, the model used to perform reasoning is described as a "first ML sub-model." On the first network device side, the model used to perform reasoning is described as a "target ML sub-model." The ML model includes a first ML sub-model and a target ML sub-model. Here, the reasoning result obtained based on the "first ML sub-model" is described as a "first reasoning result." The reasoning result obtained based on the "target ML sub-model" is described as a "target reasoning result." The target reasoning result is the final reasoning result of the ML model. Among them, the first network device can be the above-mentioned access network device, core network device or network control device.

[0189] The embodiment of the present application provides a first collaborative reasoning method, which is applied in the machine learning process. Figure 4 , the collaborative reasoning method includes the following steps:

[0190] S400: The terminal device and the first network device respectively perform the process of "configuring a first computing radio bearer (CRB)".

[0191] The first CRB is a dedicated wireless bearer used to implement the orderly transmission, encryption and decryption, and duplication detection of information related to the reasoning operation. In other words, the terminal device and the first network device transmit the information related to the reasoning operation through the first CRB. Here, the information related to the reasoning operation can be, for example, but not limited to Figure 4 The information shown is: reasoning requirement information, information of the first ML sub-model, first reasoning result, target reasoning result. It should be noted that the first network device at this time is an access network device. Figure 5 A possible configuration process of the first CRB is shown:

[0192] S400a: The first network device determines configuration information of a first CRB.

[0193] The configuration information of the first CRB may include the following information:

[0194] Item 1: The identifier of the first CRB. The identifier of the first CRB uniquely identifies a CRB.

[0195] Item 2: The serial number size of the first CRB. The serial number size of the first CRB indicates the length of the carrier for transmitting inference-related information (such as information of the first ML sub-model, the first inference result, and the target inference result). The serial number size of the first CRB can be 12 bits, 18 bits, etc. The embodiment of the present application does not limit the serial number size of the first CRB.

[0196] Item 3: The discard time of the first CRB. The discard time of the first CRB indicates that the terminal device discards or releases the first CRB after a certain period of time. For example, if the discard time of the first CRB is "5 minutes", it indicates that the terminal device keeps the first CRB for 5 minutes. After 5 minutes, the terminal device discards or releases the first CRB.

[0197] Item 4: Header compression information of the first CRB. The header compression information of the first CRB indicates the compression information of the first CRB. For example, the header compression information is the value of the maximum context identifier. In this case, the information of the first ML sub-model (or the first inference result, or the target inference result) is first compressed according to the value of the maximum context identifier, and then the compressed result is transmitted through the first CRB.

[0198] It should be noted that, among the above four information items, the configuration information of the first CRB includes the identifier of the first CRB to uniquely identify a CRB. Optionally, the configuration information of the first CRB includes at least one of the sequence number size of the first CRB, the discard time of the first CRB, and the header compression information of the first CRB.

[0199] S400b: The first network device sends configuration information of the first CRB to the terminal device. Correspondingly, the terminal device receives the configuration information of the first CRB from the first network device.

[0200] S400c. The terminal device configures the first CRB according to the configuration information of the first CRB.

[0201] In this way, when the terminal device obtains the configuration information of the first CRB, the terminal device can configure the first CRB to transmit reasoning-related information through the first CRB.

[0202] It should be noted that S400 is an optional step. Here, when the PDCP layer is associated with the CRB, the collaborative reasoning method of the embodiment of the present application includes S400, that is, executing the process of "configuring the first CRB". When the PDCP layer is not associated with the CRB, the collaborative reasoning method of the embodiment of the present application does not include S400, that is, there is no need to execute the process of "configuring the first CRB".

[0203] S401: A terminal device sends reasoning requirement information to a first network device. Correspondingly, the first network device receives the reasoning requirement information from the terminal device.

[0204] The inference requirement information includes the time information for the terminal device to obtain the target inference result. Here, the time information can be specifically implemented as "time period information," such as the time period from a first time point to a second time point. The first time point can be the time point when the terminal device executes S401. The second time point can be the latest time point at which the terminal device obtains the target inference result. Alternatively, the first time point is t1. The second time point is t2. t1 and t2 can be any pre-specified time points. In other words, the terminal device must obtain the target inference result within the "time period indicated by the above time information." The inference requirement information also includes the full ML model information or the identifier of the ML model. When the "inference requirement information includes the full ML model information," the first network device does not need to store the ML model, thereby reducing the storage space required by the first network device. Here, the full ML model information refers to information that fully describes the ML model, such as the source code describing the ML model, the executable program code of the ML model, or the partially or fully compiled code of the ML model.

[0205] Optionally, the inference requirement information also includes at least one of the following: the input size of the ML model and the computing power information of the terminal device. The input size of the ML model indicates the amount of input data used for ML inference, for example, it can be expressed in bytes. The computing power information of the terminal device can also be described as the computing power of the terminal device, which can be understood as the ability to indicate or evaluate the speed at which the terminal device processes data. For example, it is the data output speed when the terminal device calculates a hash function, specifically expressed in FLOPS. The computing power of the terminal device is positively correlated with the speed of data processing. For example, the greater the computing power, the faster the data processing speed. Therefore, generally, the faster the terminal device performs ML model inference. The computing power of the terminal device is related to factors such as the terminal device's hardware configuration performance and the smoothness of the operating system.

[0206] S402: The first network device determines a first ML sub-model according to the inference requirement information.

[0207] For example, if the inference requirement information includes an ML model identifier, the first network device can determine the corresponding ML model based on the ML model identifier, and thus determine which model to segment. If the inference requirement information includes full ML model information, the first network device can segment the ML model included in the inference requirement information.

[0208] In the case where the ML model sets a segmentation option, the first network device determines the segmentation option corresponding to the first ML sub-model according to the inference requirement information.

[0209] Next, we will first introduce the ML model splitting options: Splitting options refer to the options defined between adjacent layers in the ML model, which are used to split the ML model. A splitting option corresponds to a split position in the ML model. For example, see Figure 1 , Figure 1 A schematic diagram of the structure of an ML model is shown. Figure 1 In , the ML model split options are represented by numbers, such as 0, 1, 2, 3, etc. Figure 1 In the example, the split option "0" represents the option between the input layer and the first hidden layer of the ML model. The split position corresponding to the split option "0" is as follows: Figure 1 The dotted line between the input layer and the first hidden layer in the figure is shown. If the split option corresponding to the first ML sub-model is "0", it means that the first ML sub-model includes the input layer of the ML model, and the terminal device needs to process the input data at the input layer. The split option "1" indicates the option between the first hidden layer and the second hidden layer of the ML model. The split position corresponding to the split option "1" is shown as follows: Figure 1The dotted line between the first hidden layer and the second hidden layer is shown in Figure 1. If the split option corresponding to the first ML sub-model is "1", it means that the first ML sub-model includes the input layer and the first hidden layer of the ML model, and the terminal device needs to process the input data at the input layer and the first hidden layer. Split option "2" represents the option between the second hidden layer and the third hidden layer of the ML model. The split position corresponding to split option "2" is shown in Figure 1. Figure 1 The dotted line between the second hidden layer and the third hidden layer in the ML model is shown. If the split option corresponding to the first ML sub-model is "2", it means that the first ML sub-model includes the input layer, the first hidden layer and the second hidden layer of the ML model, and the terminal device needs to process the input data at the input layer, the first hidden layer and the second hidden layer. The split option "3" indicates the option between the third hidden layer and the output layer of the ML model. The split position corresponding to the split option "3" is as follows Figure 1 The dashed line between the third hidden layer and the output layer in the ML model is shown. If the split option corresponding to the first ML sub-model is "3", it means that the first ML sub-model includes the input layer, the first hidden layer, the second hidden layer, and the third hidden layer of the ML model. The terminal device needs to process the input data according to the input layer, the first hidden layer, the second hidden layer, and the third hidden layer. If other split options are used in the ML model, the meaning of the other split options can be deduced similarly.

[0210] Then, still Figure 1 Taking the ML model shown in the figure as an example, if the first network device selects split option "2", the first ML sub-model includes the input layer, the first hidden layer, and the second hidden layer of the ML model, but does not include the third hidden layer and the output layer of the ML model. Here, the first network device performs calculations to obtain the following information:

[0211] The first item is the duration of the terminal device performing local reasoning. For example, the first network device determines the duration of the terminal device performing local reasoning based on the computing capability of the terminal device.

[0212] Item 2: Duration of time for the terminal device to send the first inference result. For example, the first network device determines the duration of time for the terminal device to send the first inference result based on the size of the first inference result and the uplink bandwidth of the terminal device.

[0213] Item 3: Duration of time for the first network device to perform local reasoning. For example, the first network device determines the duration of time for the first network device to perform local reasoning based on its own computing capability.

[0214] Item 4: Duration of time for the first network device to send the target inference result. For example, the first network device determines the duration of time for the first network device to send the target inference result based on the target inference result and the downlink bandwidth of the terminal device.

[0215] If the sum of the above durations does not exceed the time period indicated by the time information in the inference requirement information, the first network device uses segmentation option "2" as the segmentation option corresponding to the first ML sub-model. If it does, the first network device calculates whether segmentation option "1" exceeds the time period indicated by the time information in the inference requirement information. The first network device repeats the above process until the first network device determines the segmentation option corresponding to the first ML sub-model, or the first network device traverses the segmentation options of the ML model. Here, the first network device determines the segmentation option, and accordingly determines the first ML sub-model.

[0216] In addition, the above-mentioned first information (i.e., “the duration of the terminal device performing local inference”) and the second information (i.e., “the duration of the terminal device sending the first inference result”) can also be calculated by the terminal device and reported by the terminal device to the first network device. In this case, the first network device can determine the third information (i.e., “the duration of the first network device performing local inference”) and the fourth information (i.e., “the duration of the first network device sending the target inference result”), so that the first network device can determine the segmentation options corresponding to the first ML sub-model. For details, please refer to the relevant instructions in the previous paragraph, which will not be repeated here. The specific instructions for “the terminal device determines the above-mentioned first information” and “the terminal device determines the above-mentioned second information” are as follows:

[0217] Taking "the terminal device determines the first information above" as an example, when the terminal device knows the "computational load of each layer of the ML model", the terminal device determines the duration of the terminal device's local inference execution based on its own computing power and the "computational load of each layer of the ML model". Figure 1 Taking the ML model shown as an example, when the terminal device obtains the "computational load of the input layer of the ML model", the terminal device calculates the "duration of the terminal device executing inference on the input layer of the ML model". When the terminal device obtains the "computational load of the input layer and the first hidden layer of the ML model", the terminal device calculates the "duration of the terminal device executing inference on the input layer and the first hidden layer of the ML model". When the terminal device obtains the "computational load of the input layer, the first hidden layer, and the second hidden layer of the ML model", the terminal device calculates the "duration of the terminal device executing inference on the input layer, the first hidden layer, and the second hidden layer of the ML model". In other words, when the terminal device traverses the segmentation options of the ML model, the first item of information also includes the "duration of the terminal device executing local inference under different segmentation options of the ML model."

[0218] Taking the example of "the terminal device determines the second information above", when the terminal device knows the "size of the inference results of each layer of the ML model", the terminal device determines the "duration for the terminal device to send the first inference result" based on the uplink bandwidth and the "size of the inference results of each layer of the ML model". Figure 1 Taking the ML model shown as an example, when the terminal device obtains the "size of the inference result of the input layer of the ML model", the terminal device calculates the "duration of the terminal device sending the inference result of the input layer of the ML model". When the terminal device obtains the "size of the inference result of the first hidden layer of the ML model", the terminal device calculates the "duration of the terminal device sending the inference result of the first hidden layer of the ML model". When the terminal device obtains the "size of the inference result of the second hidden layer of the ML model", the terminal device calculates the "duration of the terminal device sending the inference result of the second hidden layer of the ML model". In other words, when the terminal device traverses the segmentation options of the ML model, the second information also includes the "duration of the terminal device sending the first inference result under different segmentation options of the ML model". Furthermore, when the first network device selects the segmentation option corresponding to the first ML sub-model, the "duration of the terminal device sending the first inference result" can be obtained.

[0219] Here, the above-mentioned first information and second information can be carried in the same message as the inference requirement information, or can be carried in different messages, and the embodiment of the present application does not limit this.

[0220] It should be noted that the first ML sub-model is a part of the ML model. The first ML sub-model includes at least the input layer of the ML model. That is, the terminal device at least performs the processing of the input layer to avoid providing input data to the first network device and prevent data privacy exposure. Here, Figure 1 Taking the ML model shown in the figure as an example, the minimum value of the split option corresponding to the first ML sub-model is "0." Furthermore, after the first network device splits the ML model and determines the first ML sub-model, it also determines the target ML sub-model. That is, the output data of the first ML sub-model corresponds to the input data of the target ML sub-model.

[0221] If the ML model does not have a split option set, the first network device autonomously determines the split location of the ML model and splits the ML model to obtain two ML sub-models. Here, the model used by the terminal device for inference is recorded as "ML sub-model a", and the model used by the first network device for inference is recorded as "ML sub-model b". Afterwards, the first network device determines the above four pieces of information (i.e., the duration of the terminal device performing local inference, the duration of the terminal device sending the first inference result, the duration of the first network device performing local inference, and the duration of the first network device sending the target inference result). If the sum of the above durations does not exceed the time period indicated by the time information in the inference requirement information, the first network device uses "ML sub-model a" as the first ML sub-model. Correspondingly, "ML sub-model b" is used as the target ML sub-model. If it does, the first network device re-determines the split location and repeats the above process until the first network device determines the first ML sub-model, or the number of times the first network device repeatedly determines the split location meets the preset value.

[0222] S403: The first network device sends information about the first ML sub-model to the terminal device. Correspondingly, the terminal device receives the information about the first ML sub-model from the first network device.

[0223] The first ML sub-model is used by the terminal device to perform an inference operation to obtain a first inference result. For example, if the first network device selects split option "1," the first ML sub-model includes the input layer and the first hidden layer of the ML model, but does not include the second hidden layer, the third hidden layer, and the output layer of the ML model.

[0224] Here, see Figure 6 The specific implementation of S403 is described through two possible implementation methods:

[0225] As a first possible implementation, when the ML model of the first network device is synchronized with that of the terminal device, the first network device indicates the first ML sub-model to the terminal device through indication information, specifically as follows: Figure 6 As shown in the "First Possible Implementation" block diagram in

[15] , "synchronizing the ML model between the first network device and the terminal device" means that the meaning represented by the ML model segmentation options applies to both the first network device and the terminal device. In other words, both the first network device and the terminal device have the same understanding of the meaning represented by the ML model segmentation options. S403 is specifically implemented as S403b. Figure 6 Detailed descriptions of the steps shown are as follows:

[0226] S403a: The first network device sends model information 1 to the terminal device. Correspondingly, the terminal device receives the model information 1 from the first network device.

[0227] The model information 1 indicates the corresponding relationship between the first candidate indication information and the first segmentation position. The first segmentation position refers to the segmentation position when segmenting the ML model.

[0228] For example, the ML model is segmented by layer, and the meanings of different segmentation options are defined, as follows: Figure 1 As shown. Here, a first candidate indication information is implemented as a segmentation option, and different first candidate indication information is implemented as different segmentation options. The first segmentation position is the segmentation position corresponding to the segmentation option. If the first target indication information is implemented as segmentation option "1", it means that the segmentation is performed between the first hidden layer and the second hidden layer of the ML model. In this way, the first ML sub-model includes the input layer and the first hidden layer of the ML model, and the target ML sub-model includes the second hidden layer, the third hidden layer and the output layer of the ML model.

[0229] Optionally, in the scenario of a single ML model, the model information 1 may not carry the identifier of the ML model. In the scenario of multiple ML models, the model information 1 carries the identifier of the ML model, so that the terminal device determines the corresponding model based on the identifier of the ML model. Exemplarily, in the scenario of multiple ML models, the identifier of the ML model is predefined between the terminal device and the first network device, and one ML model identifier uniquely identifies one ML model. For example, the identifier 1 of the ML model represents the Alexnet model, the identifier 2 of the ML model represents the 16-layer visual geometry group 16 (VGG16) model, and the identifier 3 of the ML model represents the ResNet-152 model. For another example, the identifiers of the ML models are AlexNet, VGG16, ResNet-152, etc.

[0230] It should be noted that step S403a is optional. For example, if the terminal device and the first network device have previously obtained model information 1 from other network devices, then step S403a need not be performed. Alternatively, the first network device and the terminal device can obtain model information 1 from a network control device to achieve model synchronization between the first network device and the terminal device. The network control device may be an OAM device.

[0231] S403b: The first network device sends first target indication information to the terminal device. Correspondingly, the terminal device receives the first target indication information from the first network device.

[0232] The first target indication information indicates the segmentation position of the ML model. Here, the first target indication information includes the segmentation option corresponding to the first ML sub-model, and the segmentation option is used to indicate the segmentation position of the ML model, so that the terminal device obtains the first ML sub-model by segmenting the ML model. Optionally, in the scenario of a single ML model, the first target indication information may not carry the identifier of the first ML sub-model. In the scenario of multiple ML models, the first target indication information carries the identifier of the first ML sub-model. Here, the identifier of the first ML sub-model is the same as the identifier of the ML model.

[0233] For example, Figure 1 Taking the illustrated scenario as an example, when the first network device determines that the segmentation option is "1," the first target indication information includes the segmentation option as "1." Accordingly, the first ML sub-model includes the input layer and the first hidden layer of the ML model, and the terminal device needs to process the input data at the input layer and the first hidden layer.

[0234] It should be noted that when the first network device needs to execute S403a and S403b, the first network device may execute S403a first and then execute S403b, or the first network device may execute S403a and S403b simultaneously. In addition, the model information 1 and the first target indication information may also be carried in the same message, that is, the first network device sends the "segmentation options corresponding to the first ML sub-model" and the meaning represented by the "segmentation options corresponding to the first ML sub-model" to the terminal device, which is not limited in this embodiment of the present application.

[0235] S403c: The terminal device determines a first ML sub-model according to the model information 1 and the first target indication information.

[0236] For example, in a scenario with multiple ML models, when the terminal device obtains model information 1, it can obtain the ML model segmentation method corresponding to the ML model identifier. Under the segmentation method of "segmentation by layer" indicated by model information 1, the terminal device can then determine which model to segment and "which layers in the ML model to be segmented belong to the first ML sub-model" in combination with the first target indication information, thereby obtaining the first ML sub-model. For example, when the first target indication information includes a segmentation option of "1", the terminal device segments the ML model, i.e., segments between the first hidden layer and the second hidden layer, to obtain the first ML sub-model.

[0237] In this way, when the ML model between the first network device and the terminal device is synchronized, the first network device can send the first target indication information (i.e., the segmentation option corresponding to the first ML sub-model to indicate the segmentation position of the ML model) to the terminal device, so that the terminal device obtains the first ML sub-model, thereby saving transmission resources.

[0238] As a second possible implementation, when the inference requirement information includes the full information of the ML model, see Figure 6 In the block diagram of "Second Possible Implementation", S403 is specifically implemented as S403a:

[0239] S403a: The first network device sends the first ML sub-model full information to the terminal device. Correspondingly, the terminal device receives the first ML sub-model full information from the first network device.

[0240] Specifically, the full information of the first ML sub-model refers to information that can fully describe the first ML sub-model, such as source code describing the first ML sub-model, executable program code of the first ML sub-model, or partially or fully compiled code of the first ML sub-model. In this way, even if model synchronization is not performed between the first network device and the terminal device, the terminal device can still obtain the first ML sub-model.

[0241] S404: The terminal device calculates a first inference result according to the first ML sub-model.

[0242] The first ML sub-model includes at least the input layer of the ML model. For example, taking "the first ML sub-model includes the input layer and the first hidden layer of the ML model" as an example, the first inference result is the inference result of the first hidden layer.

[0243] Here, the terminal device inputs the data into the first ML sub-model, and uses the first ML sub-model to calculate the input data to obtain a first inference result. The input data is the input data of the first ML sub-model generated by the terminal device, that is, the input data is generated by the terminal device and serves as the input data of the first ML sub-model. For example, in the scenario of terminal device transmit power self-optimization, the terminal device can optimize its own transmit power through the power ML model. Specifically, the terminal device obtains the first power ML sub-model, and uses the transmit power at the current moment or at a certain moment(s) before as the input data of the first power ML sub-model. The terminal device uses the first power ML sub-model to perform inference calculation on this(some) transmit power value(s) to obtain a first inference result. It can be seen that the terminal device does not need to provide the input data of the ML model to the network device, reducing the risk of "data privacy exposure".

[0244] S405: The terminal device sends the first inference result to the first network device. Correspondingly, the first network device receives the first inference result from the terminal device.

[0245] Here, the first inference result refers to the complete first inference result. Figure 1Taking “the first ML sub-model includes the input layer and the first hidden layer of the ML model” as an example, the first inference result includes the inference result of the first hidden layer.

[0246] S406: The first network device calculates a target inference result based on the first inference result and the target ML sub-model.

[0247] The target ML sub-model at least includes the output layer of the ML model. Here, the input data of the target ML sub-model corresponds to the output data of the first ML sub-model. For example, Figure 1 Taking "the first ML sub-model includes the input layer and the first hidden layer of the ML model" as an example, the target ML sub-model includes the second hidden layer, the third hidden layer, and the output layer of the ML model.

[0248] Among them, the target inference result is the final inference result of the ML model.

[0249] Exemplarily, the first network device inputs the first inference result into the target ML sub-model, and uses the target ML sub-model to perform processing of the second hidden layer, the third hidden layer, and the output layer to obtain a target inference result.

[0250] Taking the above-mentioned terminal transmission power self-optimization scenario as an example, the first network device uses the first inference result obtained by the terminal device using the first power ML sub-model as the input data of the target power ML sub-model, and uses the target power ML sub-model to perform inference calculation to obtain the target inference result, that is, the optimized terminal device transmission power.

[0251] S407: The first network device sends the target inference result to the terminal device. Correspondingly, the terminal device receives the target inference result from the first network device.

[0252] Taking the above-mentioned terminal device transmission power self-optimization scenario as an example, after the terminal device receives the optimized terminal device transmission power sent by the first network device, it can use the optimized transmission power to send data.

[0253] It should be noted that in the above-mentioned interaction steps between the terminal device and the first network device (such as S401, S403, S405 and S407), the terminal device and the first network device can send relevant information about the reasoning operation based on the existing protocol stack. For example, the relevant information about the reasoning operation is carried in an RRC message or a non-access stratum (NAS) message. The terminal device and the first network device can also send relevant information about the reasoning operation based on a new protocol stack.

[0254] For example, when the first network device is implemented as an access network device, a dedicated protocol (such as the data analytics protocol (DAP)) can be used between the terminal device and the access network device to send information related to the reasoning operation to achieve functions such as segmentation, sorting, integrity protection, encryption and decryption of the relevant information. Here, the PDCP layer is associated with a dedicated radio bearer (such as CRB) to achieve orderly transmission, encryption and decryption, and duplication detection of information related to the reasoning operation. Figure 7a , Figure 7a The protocol stack between the terminal device and the access network device is shown. The protocol stack is used to transmit information related to the inference operation between the terminal device and the access network device. The protocol stack mainly includes the DAP layer, the PDCP layer, the RLC layer, the MAC layer, and the PHY layer. Among them, the DAP layer, the PDCP layer, the RLC layer, the MAC layer, and the PHY layer all belong to the access stratum (AS) layer. Here, the information related to the inference operation can be, for example, but not limited to, the following information: inference requirement information, information of the first ML sub-model, the first inference result, and the target inference result.

[0255] For another example, when the first network device is implemented as a core network device, a dedicated protocol (such as high data analytics protocol (HDAP)) can be used between the terminal device and the core network device to send relevant information to achieve functions such as segmentation, sorting, integrity protection, encryption and decryption of relevant information. Figure 7b , Figure 7b The protocol stack between the terminal device and the core network device is shown. Similarly, the protocol stack is used to transmit information related to reasoning operations between the terminal device and the core network device. The protocol stack mainly includes the HDAP layer. It should be noted that Figure 7b The protocol stack for interaction between access network equipment and core network equipment is omitted. For an introduction to the protocol stack for interaction between terminal equipment and access network equipment, please refer to Figure 7a The relevant instructions will not be repeated here.

[0256] In addition, in the first collaborative reasoning method provided in the embodiment of the present application, S400 can be executed before any one of steps S401 to S407, or can be executed simultaneously with any one of steps S401 to S407, and the embodiment of the present application does not limit this. When S400 is executed simultaneously with a certain step, the "configuration information of the first CRB" and the information transmitted by the step can be carried in the same message or in different messages, and the embodiment of the present application does not limit this. For example, taking the simultaneous execution of S400 and S403 as an example, the "configuration information of the first CRB" and the "first ML sub-model" can be carried in the same message or in different messages.

[0257] In the collaborative reasoning method provided in the embodiment of the present application, the terminal device uses the first ML sub-model to perform partial reasoning operations to obtain a first reasoning result. After the terminal device sends the first reasoning result, the first network device calculates all the information of the first reasoning result in combination with the target ML sub-model to obtain the target reasoning result, and then provides it to the terminal device, so that the terminal device does not need to perform a complete reasoning operation, thereby reducing the delay for the terminal device to obtain the target reasoning result. Furthermore, what the terminal device provides to the network device is the intermediate result of the ML model calculation rather than the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0258] exist Figure 4 In the communication process shown, if the terminal device undergoes a handover, RRC connection recovery, or RRC connection re-establishment, thereby connecting to the second network device, the terminal device receives the target reasoning result from the second network device. Here, taking the "handover" of the terminal device as an example, after the first network device obtains the information provided by the terminal device (such as the complete first reasoning result), the first network device determines that a handover needs to be initiated for the terminal device, and the first network device does not perform the reasoning operation. Alternatively, after the first network device obtains the information provided by the terminal device, the first network device determines that a handover needs to be initiated for the terminal device, and the computing power of the second network device is better than that of the first network device, the first network device may not perform the reasoning operation, and the second network device performs the reasoning operation. Taking the "RRC connection recovery" or "RRC connection re-establishment" of the terminal device as an example, after the first network device obtains the information provided by the terminal device (such as the complete first reasoning result), if the first network device receives a retrieve UE context request message from the second network device, the first network device does not perform the reasoning operation, and the second network device performs the reasoning operation. Here, if the first network device receives a request message for retrieving the terminal device context from the second network device, it indicates that the terminal device is connected to the second network device. In the scenario where the first network device does not perform inference operations, similarly, the ML model includes a first ML sub-model and a target ML sub-model. On the terminal device side, the model used to perform inference is described as a "first ML sub-model", and the obtained inference result is described as a "first inference result". On the second network device side, the model used to perform inference is described as a "target ML sub-model", and the obtained inference result is described as a "target inference result". The second network device can be the above-mentioned access network device, core network device or network control device. Optionally, in the case of transmitting inference-related information through CRB, the CRB between the terminal device and the first network device is described as a "first CRB", and the CRB between the terminal device and the second network device is described as a "target CRB".

[0259] Below, taking the switching of a terminal device (i.e., the terminal device switches from a first network device to a second network device, at this time, the first network device is a first access network device, and the second network device is a second access network device) as an example, the second collaborative reasoning method provided in the embodiment of the present application is introduced. This collaborative reasoning method is applied in the machine learning process. Figure 8 The collaborative reasoning method may include S400 to S404, and the following steps:

[0260] S800: The terminal device and the second network device respectively perform the process of "configuring the target CRB".

[0261] The target CRB is also a dedicated wireless bearer for orderly sending, encryption and decryption, and repeatability detection of information related to the reasoning operation. In other words, the terminal device and the second network device transmit information related to the reasoning operation through the target CRB. Here, the information related to the reasoning operation can be, for example, but not limited to Figure 8 The information shown is: the second part of the first reasoning result, the entire information of the first reasoning result, and the target reasoning result. Figure 9a A possible configuration process of a target CRB is shown:

[0262] Optionally, if a first CRB exists between the terminal device and the first network device, execute S800a:

[0263] S800a. The first network device sends configuration information of the first CRB to the second network device.

[0264] Among them, the relevant description of "configuration information of the first CRB" can be found in the introduction of S400a, which will not be repeated here.

[0265] For example, in a handover scenario, the configuration information of the first CRB may be carried in a handover request message. Of course, the configuration information of the first CRB may also be carried in other messages, which is not limited in the embodiment of the present application.

[0266] It should be noted that S800a is an optional step. Here, if a first CRB exists between the terminal device and the first network device, the first network device may or may not execute S800a. If no first CRB exists between the terminal device and the first network device, the first network device does not need to execute S800a.

[0267] S800b: The second network device determines configuration information of the target CRB.

[0268] The configuration information of the target CRB may include the following information:

[0269] The first item is the identifier of the target CRB. The identifier of the target CRB uniquely identifies a CRB.

[0270] Item 2: Target CRB Sequence Number Size. The target CRB sequence number size indicates the length of the carrier for transmitting inference-related information (such as information about the target ML sub-model, all information about the first inference result, and the target inference result). The target CRB sequence number size can be 12 bits, 18 bits, etc., and the embodiment of the present application does not limit the target CRB sequence number size.

[0271] Item 3: Target CRB discard time. The target CRB discard time indicates that the terminal device discards or releases the target CRB after a certain period of time. For example, if the target CRB discard time is "5 minutes," it indicates that the terminal device retains the target CRB for 5 minutes. After 5 minutes, the terminal device discards or releases the target CRB.

[0272] Item 4: Header compression information of the target CRB. The header compression information of the target CRB indicates the compression information of the target CRB. For example, the header compression information is the value of the maximum context identifier. In this case, the information of the first ML sub-model (or the first inference result, or the target inference result) is first compressed according to the value of the maximum context identifier, and then the compressed result is transmitted through the target CRB.

[0273] It should be noted that, among the above four items of information, the configuration information of the target CRB includes the identifier of the target CRB to uniquely identify a CRB. Optionally, the configuration information of the target CRB includes at least one of the sequence number size of the target CRB, the discard time of the target CRB, and the header compression information of the target CRB. S800a is an optional step. In the case of executing S800a, the second network device determines the configuration information of the target CRB based on the configuration information of the first CRB. For example, the second network device modifies some parameters in the configuration information of the first CRB to obtain the configuration information of the target CRB. In the case where S800a is not executed, the second network device may also determine the configuration information of the target CRB without combining the configuration information of the first CRB.

[0274] S800c: The second network device sends the configuration information of the target CRB to the first network device. Correspondingly, the first network device receives the configuration information of the target CRB from the second network device.

[0275] For example, in a handover scenario, the configuration information of the target CRB is carried in a handover request acknowledgement message. The handover request acknowledgement message is a message sent to the first network device after the second network device has completed the handover preparation process. Of course, the configuration information of the target CRB can also be carried in other messages, and this embodiment of the present application is not limited to this.

[0276] S800d: The first network device sends the configuration information of the target CRB to the terminal device. Correspondingly, the terminal device receives the configuration information of the target CRB from the first network device.

[0277] S800e: The terminal device configures the target CRB according to the configuration information of the target CRB.

[0278] Exemplarily, when the terminal device is configured with the first CRB, the terminal device modifies the first CRB according to the configuration information of the target CRB to obtain the target CRB. When the terminal device is not configured with the first CRB, the terminal device configures the target CRB according to the configuration information of the target CRB.

[0279] Here, after the terminal device completes the configuration of the target CRB, optionally, the terminal device sends a configuration confirmation to the second network device. Correspondingly, the second network device receives the reconfiguration confirmation from the terminal device.

[0280] In this way, in a scenario where a terminal device switches, after the second network device determines the configuration information of the target CRB, the second network device provides the configuration information of the target CRB to the terminal device through the first network device, so that the terminal device configures the target CRB. Furthermore, inference-related information can be transmitted between the terminal device and the second network device via the target CRB.

[0281] It should be noted that S800 is an optional step. Here, when the PDCP layer is associated with the CRB, the collaborative reasoning method of the embodiment of the present application includes S800, that is, the process of performing "configuring the target CRB". When the PDCP layer is not associated with the CRB, the collaborative reasoning method of the embodiment of the present application does not include S800, that is, there is no need to perform the process of "configuring the target CRB".

[0282] S801: A first network device sends information about a target ML sub-model to a second network device. Correspondingly, the second network device receives the information about the target ML sub-model from the first network device.

[0283] The input data of the target ML sub-model corresponds to the output data of the first ML sub-model. Here, after the first network device executes S402, the target ML sub-model can be obtained.

[0284] For example, the specific implementation process of S801 is described below through Example 1 and Example 2:

[0285] Example 1: In the case of ML model synchronization between the first network device and the second network device, the first network device indicates the target ML sub-model to the second network device through the second target indication information, specifically as follows: Figure 9b As shown in the "Example 1" block diagram. Here, "ML model synchronization between the first network device and the second network device" means that the meaning represented by the ML model segmentation options applies to both the first network device and the second network device. In other words, the first network device and the second network device have the same understanding of the meaning represented by the ML model segmentation options. S801 is specifically implemented as S801c. Figure 9b Detailed descriptions of the steps shown are as follows:

[0286] S801a: The first network device sends an ML model query request to the second network device. Correspondingly, the second network device receives the ML model query request from the first network device.

[0287] The ML model query request is used to request the ML models supported by the second network device and the segmentation mode of the ML model supported by the second network device. Here, when the segmentation mode of the ML model supported by the second network device is "segment by layer", the meaning of different segmentation options can be found in Figure 1 The relevant instructions will not be repeated here.

[0288] S801b: The second network device sends model information 2 to the first network device. Correspondingly, the first network device receives model information 2 from the second network device.

[0289] The model information 2 indicates the corresponding relationship between the second candidate indication information and the second segmentation position. The second segmentation position refers to the segmentation position when segmenting the ML model.

[0290] For example, the ML model is segmented by layer, and the meanings of different segmentation options are defined, as follows: Figure 1 As shown. Here, one second candidate indication information is implemented as a segmentation option, and different second candidate indication information is implemented as different segmentation options. The second segmentation position is the segmentation position corresponding to the segmentation option. If the second target indication information is implemented as segmentation option "1", it means that the segmentation is performed between the first hidden layer and the second hidden layer of the ML model. In this way, the first ML sub-model includes the input layer and the first hidden layer of the ML model, and the target ML sub-model includes the second hidden layer, the third hidden layer and the output layer of the ML model.

[0291] Optionally, in a scenario with a single ML model, the model information 2 may not carry the identifier of the ML model. In a scenario with multiple ML models, the model information 2 carries the identifier of the ML model, so that the first network device determines the corresponding model according to the identifier of the ML model.

[0292] It should be noted that S801a and S801b are optional steps. For example, if the first network device and the second network device obtain model information 2 from other network devices in advance, there is no need to execute S801a and S801b. Here, the first network device and the second network device may also obtain model information 2 from the network control device to achieve model synchronization between the first network device and the second network device. The network control device may be an OAM device. Furthermore, when executing S801a and S801b, the second network device may execute S801b without executing S801a, that is, the second network device can directly feedback model information 2 to the first network device. Of course, the second network device may also execute S801a and S801b, that is, the second network device will only feedback model information 2 to the first network device when the first network device requests it from the second network device.

[0293] S801c: The first network device sends second target indication information to the second network device. Correspondingly, the second network device receives the second target indication information from the first network device.

[0294] The second target indication information indicates the segmentation position of the ML model. Here, the second target indication information includes a segmentation option corresponding to the target ML sub-model, and indicates the segmentation position of the ML model through the segmentation option, so that the second network device obtains the target ML sub-model by segmenting the ML model. Exemplarily, in a switching scenario, the second target indication information can be carried in a switching request message. The switching request message is used to request that the terminal device be switched to the second network device. After the second network device has completed the switching preparation process, the second network device sends a switching request confirmation message to the first network device.

[0295] Optionally, in a scenario with a single ML model, the second target indication information may not carry the identifier of the target ML sub-model. In a scenario with multiple ML models, the second target indication information carries the identifier of the first ML sub-model. Here, the identifier of the target ML sub-model is the same as the identifier of the ML model.

[0296] For example, Figure 1Taking the illustrated scenario as an example, when the first network device determines that the split option is "1," the second target indication information includes the split option as "1." Accordingly, the first ML sub-model includes the input layer and the first hidden layer of the ML model, and the target ML sub-model includes the second hidden layer, the third hidden layer, and the output layer of the ML model. In this case, the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0297] S801d: The second network device determines a target ML sub-model according to the model information 2 and the second target indication information.

[0298] For example, in a scenario with multiple ML models, when the second network device obtains model information 2, it can obtain the ML model segmentation method corresponding to the ML model identifier. Under the segmentation method of "segmentation by layer" indicated by model information 2, the second network device, combined with the second target indication information, can determine which model to segment and "which layers in the ML model to be segmented belong to the target ML sub-model," thereby obtaining the target ML sub-model. For example, when the second target indication information includes a segmentation option of "1," the second network device segments the ML model, i.e., segments between the first hidden layer and the second hidden layer, to obtain the target ML sub-model.

[0299] In this way, when the ML models between the first network device and the second network device are synchronized, the first network device can send the second target indication information (i.e., the segmentation option corresponding to the target ML sub-model to indicate the segmentation position of the ML model) to the second network device, so that the second network device can obtain the target ML sub-model, thereby saving transmission resources.

[0300] Example 2: When the reasoning requirement information includes the full information of the ML model, such as Figure 9b As shown in the block diagram of "Example 2", S801 is specifically implemented as S801a:

[0301] S801a: The first network device sends the full information of the target ML sub-model to the second network device. Correspondingly, the second network device receives the full information of the target ML sub-model from the first network device.

[0302] Here, the full target ML submodel information refers to information that fully describes the target ML submodel, such as source code describing the target ML submodel, executable program code of the target ML submodel, or partially or fully compiled code of the target ML submodel. Thus, when a first network device provides the full target ML submodel information to a second network device, the second network device can obtain the target ML submodel without requiring model synchronization between the first and second network devices.

[0303] For the terminal device, the terminal device executes S404 to obtain the first inference result. Figure 8 Before the terminal device switches from the first network device to the second network device, based on the first inference result, the transmission status between the terminal device and the first network device can be divided into the following three situations:

[0304] The first case (such as Figure 8 (As shown in the "First Case" box diagram): All the information of the first reasoning result (i.e., the complete first reasoning result) is divided into two parts, i.e., all the information of the first reasoning result includes the first part of the information of the first reasoning result and the second part of the information of the first reasoning result. Among them, the first part of the information of the first reasoning result refers to the information of the first reasoning result provided by the terminal device to the first network device. The second part of the information of the first reasoning result refers to the information of the first reasoning result provided by the terminal device to the second network device. That is to say, after the terminal device sends the first part of the information of the first reasoning result to the first network device, the terminal device switches, i.e., switches from the first network device to the second network device, and the terminal device no longer interacts with the first network device, thereby sending the second part of the information of the first reasoning result to the second network device. In addition, in order to enable the network side to perform the reasoning operation of the target ML sub-model, the first network device needs to send the first part of the information of the first reasoning result to the second network device, so that the second network device performs the reasoning operation and obtains the target reasoning result. For details, please refer to the relevant descriptions of S802a to S802c in the first case:

[0305] S802a: The terminal device sends the first part of the first reasoning result to the first network device. Correspondingly, the first network device receives the first part of the first reasoning result from the terminal device.

[0306] For example, Figure 1 For example, in the example where "the first ML sub-model includes the input layer and the first hidden layer of the ML model," the first inference result is the inference result of the first hidden layer. The terminal device sends the first portion of the inference result of the first hidden layer to the first network device. In return, the first network device receives the first portion of the inference result of the first hidden layer from the terminal device.

[0307] It should be noted that the first network device may execute S801 first and then execute S802a, or may execute S802a first and then execute S801, or may execute S801 and S802a simultaneously, and this embodiment of the present application is not limited thereto. Furthermore, when the "target ML sub-model" is carried in the handover request message, the first network device executes S802a first and then executes S801.

[0308] S802b: The first network device sends the first part of the first reasoning result to the second network device. Correspondingly, the second network device receives the first part of the first reasoning result from the first network device.

[0309] It should be noted that, optionally, the first network device further sends the state information of the first CRB to the second network device. Correspondingly, the second network device receives the state information of the first CRB from the first network device.

[0310] The status information of the first CRB includes the identifier of the first CRB and the status corresponding to each CRB sequence number in the first CRB. For example, the status corresponding to a CRB sequence number is represented by the value of a bit. If the bit corresponding to a CRB sequence number is "0," it indicates that the data portion corresponding to the CRB sequence number has failed to be received. If the bit corresponding to a CRB sequence number is "1," it indicates that the data portion corresponding to the CRB sequence number has been successfully received. Conversely, if the bit corresponding to a CRB sequence number is "0," it indicates that the data portion corresponding to the CRB sequence number has been successfully received. If the bit corresponding to a CRB sequence number is "1," it indicates that the data portion corresponding to the CRB sequence number has failed to be received. In this way, based on the status information of the first CRB, the second network device can obtain the "data portion that the first network device failed to receive," and then request the terminal device to resend the "data portion that the first network device failed to receive." In this way, the terminal device can send the "data portion that the first network device failed to receive" to the second network device, ensuring that the second network device obtains all information about the first inference result.

[0311] S802c: The terminal device sends the second part of the first reasoning result to the second network device. Correspondingly, the second network device receives the second part of the first reasoning result from the terminal device.

[0312] At this point, the second network device can use the first part of the information of the first inference result obtained from the first network device and the second part of the information of the first inference result obtained from the terminal device as input data of the target ML sub-model to perform inference operation.

[0313] The second case (such as Figure 8 (As shown in the "Second Case" block diagram): After the terminal device sends the complete first inference result to the first network device, the terminal device switches, that is, switches from the first network device to the second network device. For details, see the relevant descriptions of S802a and S802b in the second case:

[0314] S802a: The terminal device sends all information about the first inference result to the first network device. That is, the terminal device sends the complete first inference result to the first network device. Correspondingly, the first network device receives all information about the first inference result from the terminal device.

[0315] For example, Figure 1 For example, in the example where "the first ML sub-model includes the input layer and the first hidden layer of the ML model," the first inference result is the inference result of the first hidden layer. The terminal device sends all information about the inference result of the first hidden layer to the first network device. Correspondingly, the first network device receives all information about the inference result of the first hidden layer from the terminal device.

[0316] It should be noted that the first network device may execute S801 first and then execute S802a, or may execute S802a first and then execute S801, or may execute S801 and S802a simultaneously, and this embodiment of the present application is not limited thereto. Furthermore, when the "target ML sub-model" is carried in the handover request message, the first network device executes S802a first and then executes S801.

[0317] S802b: The first network device sends all information about the first inference result to the second network device. Correspondingly, the second network device receives all information about the first inference result from the first network device.

[0318] At this point, the second network device can use all information of the first inference result obtained from the first network device as input data of the target ML sub-model to perform inference operations.

[0319] The third case (such as Figure 8 (As shown in the "Third Case" block diagram): After the terminal device obtains the first inference result, the terminal device has already switched, that is, switched from the first network device to the second network device. The terminal device does not provide the first inference result to the first network device, but provides the first inference result to the second network device. For details, see the relevant description of S802a in the third case:

[0320] S802a: The terminal device sends all information of the first reasoning result to the second network device. Correspondingly, the second network device receives all information of the first reasoning result from the terminal device.

[0321] For example, using the example of "the first ML sub-model includes the input layer and the first hidden layer of the ML model," the first inference result is the inference result of the first hidden layer. The terminal device sends all information about the inference result of the first hidden layer to the second network device. Correspondingly, the second network device receives all information about the inference result of the first hidden layer from the terminal device.

[0322] At this point, the second network device can use all the information of the first inference result obtained from the terminal device as input data of the target ML sub-model to perform inference operation.

[0323] In the above three cases, the second network device obtains all information of the first reasoning result in different ways and performs local reasoning, that is, the second network device executes S803:

[0324] S803: The second network device calculates a target inference result based on all information of the first inference result and the target ML sub-model.

[0325] For example, Figure 1 Taking "the first ML sub-model includes the input layer and the first hidden layer of the ML model" as an example, the first inference result is the inference result of the first hidden layer. The target ML sub-model includes the second hidden layer, the third hidden layer and the output layer. The second network device uses all the information of the first inference result as the input data of the target ML sub-model, and uses the target ML sub-model to perform inference calculations to obtain the target inference result. It should be noted that in the first case above, after the second network device executes S802b and S802c, the second network device integrates the first part of the information of the first inference result and the second part of the information of the first inference result to obtain all the information of the first inference result, that is, the complete first inference result, and then executes S803 to obtain the target inference result.

[0326] S804: The second network device sends the target inference result to the terminal device. Correspondingly, the terminal device receives the target inference result from the second network device.

[0327] The specific implementation process of S804 can refer to the relevant description of S407 and will not be repeated here.

[0328] It should be noted that, in the above steps, when there is an Xn interface between the first network device and the second network device, in the above steps, the first network device and the second network device transmit messages via the Xn interface. The first network device and the second network device can use the existing protocol stack to transmit relevant information, or they can use Figure 9cThe protocol stack shown transmits relevant information. For example, the message between the first network device and the second network device is carried in a high data analytics protocol type b (HDAPb) message. The HDAPb protocol supports functions such as computing data transmission (such as data segmentation and data sorting) between the first network device and the second network device, as well as computing data security (such as data integrity protection, data encryption, and data decryption). The HDAPb message can be carried in the XnAP message. Here, Figure 9c The figure shows a protocol stack between two access network devices (i.e., access network device 1 and access network device 2). This protocol stack is used to transmit information related to inference operations between the two access network devices. The protocol stack mainly includes the HDAP layer, the Xn application protocol (XnAP) layer, the stream control transmission protocol (SCTP) layer, the internet protocol (IP) layer, the L2 layer, and the L1 layer. The above-mentioned relevant information can be, for example, but not limited to, the following information: information about the target ML sub-model, the first part of the first inference result, and the entire information of the first inference result.

[0329] On the contrary, when there is no Xn interface between the first network device and the second network device, in the above steps, information is transmitted between the first network device and the second network device through the core network device. Taking "the first network device sends all the information of the first reasoning result to the second network device" as an example, the first network device sends all the information of the first reasoning result to the core network device through the NG interface. Correspondingly, the core network device receives all the information of the first reasoning result from the first network device. Afterwards, the core network device sends all the information of the first reasoning result to the second network device. Correspondingly, the second network device receives all the information of the first reasoning result from the core network device. Here, the first network device (or the second network device) and the core network device can use the existing protocol stack to transmit relevant information, or they can use Figure 9dThe protocol stack shown transmits relevant information. For example, messages between the first network device (or the second network device) and the core network device are carried in high data analytics protocol type a (HDAPa) messages. The HDAPa protocol supports functions such as computing data transmission (such as data segmentation and data sorting) and computing data security (such as data integrity protection, data encryption, and data decryption) between the first network device (or the second network device) and the core network device. HDAPa messages can be carried in next generation application protocol (NGAP) messages. Figure 9d The protocol stack between access network devices and core network devices is shown. This protocol stack is used to transmit information related to inference operations between the access network devices and the core network devices. The protocol stack mainly includes the HDAP layer, NGAP layer, SCTP layer, IP layer, L2 layer, and L1 layer.

[0330] Below, the second collaborative reasoning method provided by the embodiment of the present application is described for the "RRC connection recovery" or "RRC connection re-establishment" scenario. It should be noted that in this scenario, the terminal device experiences an RRC interruption, failure, or suspension in the area served by the first network device, and then enters the area served by the second network device and initiates RRC connection recovery or RRC connection re-establishment to the second network device.

[0331] It should be noted that in the RRC connection recovery scenario or the RRC connection re-establishment scenario, the configuration process of the target CRB (ie, the specific implementation process of S800) is as follows: Figure 10 Steps shown:

[0332] S1000a. The first network device sends configuration information of the first CRB to the second network device.

[0333] Among them, the relevant description of "the configuration information of the first CRB" can be found in the introduction of S800a, which will not be repeated here. In the "RRC connection recovery" scenario, the configuration information of the first CRB can be carried in the retrieve UE context response message. Of course, the configuration information of the first CRB can also be carried in other messages, which is not limited in this embodiment of the present application.

[0334] It should be noted that S1000a is an optional step. Here, if a first CRB exists between the terminal device and the first network device, the first network device may or may not execute S1000a. If no first CRB exists between the terminal device and the first network device, the first network device does not need to execute S1000a.

[0335] S1000b. The second network device determines configuration information of the target CRB.

[0336] The specific implementation process of S1000b can refer to the relevant description of S800b and will not be repeated here.

[0337] S1000c: The second network device sends the configuration information of the target CRB to the terminal device. Correspondingly, the terminal device receives the configuration information of the target CRB from the second network device.

[0338] S1000d. The terminal device configures the target CRB according to the configuration information of the target CRB.

[0339] The specific implementation process of S1000d can be found in the relevant description of S800e, which will not be repeated here.

[0340] In this way, in the scenario where the terminal device performs RRC connection recovery, after the second network device determines the configuration information of the target CRB, the second network device provides the configuration information of the target CRB to the terminal device so that the terminal device configures the target CRB and transmits reasoning-related information between the target CRB and the second network device.

[0341] In addition, in the "RRC connection recovery" scenario, the information transmission process between the terminal device and the network device may also include the following steps 1a to 1c:

[0342] Step 1a: The terminal device sends an RRC recovery request message to the second network device. Correspondingly, the second network device receives the RRC recovery request message from the terminal device.

[0343] The RRC recovery request message is used to request the restoration of the RRC connection. The RRC recovery request message includes an RRC recovery reason. Exemplarily, the RRC recovery reason is that the terminal device needs to send the first reasoning result.

[0344] Step 1b: The second network device sends a terminal device context retrieval request message to the first network device. Correspondingly, the first network device receives the terminal device context retrieval request message from the second network device.

[0345] The terminal device context retrieval request message is used to request the context of the terminal device. Exemplarily, the terminal device context retrieval request message includes an RRC recovery reason. Here, the RRC recovery reason is still that the terminal device needs to send the first reasoning result.

[0346] Step 1c: The first network device sends a retrieve UE context response message to the second network device. Correspondingly, the second network device receives the retrieve UE context response message from the first network device.

[0347] In the "RRC connection re-establishment" scenario, the information transmission process between the terminal device and the network device includes the following steps 2a to 2c:

[0348] Step 2a: The terminal device sends an RRC reestablishment request message to the second network device. Correspondingly, the second network device receives the RRC reestablishment request message from the terminal device.

[0349] The RRC re-establishment request message is used to request re-establishment of the RRC connection. The RRC re-establishment request message includes an RRC re-establishment cause. Exemplarily, the RRC re-establishment cause is that the terminal device needs to send the first reasoning result.

[0350] In step 2b, the second network device sends a request message to retrieve the terminal device context to the first network device. Correspondingly, the first network device receives the request message from the second network device. For details about step 2b, refer to the description of step 1b in the "RRC Connection Recovery" scenario and are not repeated here.

[0351] In step 2c, the first network device sends a retrieve UE context response message to the second network device. In response, the second network device receives the retrieve UE context response message from the first network device. For an introduction to step 2c, refer to the description of step 1c in the "RRC Connection Recovery" scenario and will not be repeated here.

[0352] Here, in the "RRC connection recovery" or "RRC connection re-establishment" scenario, in the specific implementation process of S801, the information of the target ML sub-model (such as the second target indication information or the full information of the target ML sub-model) can be carried in the retrieval terminal device context response message.

[0353] It should be noted that in the "RRC connection recovery" or "RRC connection re-establishment" scenario, all the information of the first reasoning result (ie, the complete first reasoning result) can still be divided into two parts, see Figure 8 The relevant descriptions are not repeated here. That is to say, after the terminal device sends the first part of the information of the first inference result to the first network device, if the first network device receives a retrieval terminal device context request message from the second network device, the first network device no longer interacts with the terminal device. In this case, the terminal device and the second network device perform an RRC connection recovery process, and the terminal device sends the second part of the information of the first inference result to the second network device. In addition, in order to enable the network side to perform the inference operation of the target ML sub-model, the first network device also sends the first part of the information of the first inference result to the second network device, so that the second network device performs the inference operation, see Figure 8 The specific implementation of the first case.

[0354] Alternatively, after the terminal device sends the complete first inference result to the first network device, if the first network device receives a request message for retrieving the terminal device context from the second network device, the first network device sends the complete first inference result to the second network device so that the second network device performs the inference operation, see Figure 8 The specific implementation of the second case.

[0355] Alternatively, the terminal device and the second network device perform an RRC connection recovery process. The first network device receives a request message to retrieve the terminal device context from the second network device, and the first network device no longer interacts with the terminal device. After the terminal device obtains the first inference result, the terminal device provides the complete first inference result to the second network device, see Figure 8 The specific implementation of the third case.

[0356] In the second collaborative reasoning method provided in an embodiment of the present application, even if the terminal device switches from the first network device to the second network device, or the terminal device performs RRC connection recovery to access the second network device, or the terminal device performs RRC connection re-establishment to access the second network device, after the terminal device obtains the first reasoning result, it can directly (such as the terminal device sends all the information of the first reasoning result to the second network device) or indirectly (such as the first network device forwards the first part of the information or all the information of the first reasoning result of the terminal device to the second network device) provide the second network device with all the information of the first reasoning result. The second network device can calculate all the information of the first reasoning result in combination with the target ML sub-model to obtain the target reasoning result, and then provide it to the terminal device, so that the terminal device does not need to perform a complete reasoning operation, reducing the delay for the terminal device to obtain the target reasoning result. Similarly, what the terminal device provides to the network device is the intermediate result of the ML model calculation instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0357] The above second collaborative reasoning method is introduced by taking the scenario of "the first network device does not perform the reasoning operation" as an example. Below, the collaborative reasoning method of the embodiment of the present application is described by taking the scenario of "the first network device performs the reasoning operation" as an example. Here, still taking the switching of the terminal device as an example, after the first network device obtains the complete first reasoning result provided by the terminal device, the first network device determines that there is no need to initiate a switch for the terminal device, then the first network device performs the reasoning operation. Taking the "RRC connection recovery" or "RRC connection re-establishment" of the terminal device as an example, after the first network device obtains the complete first reasoning result provided by the terminal device, if the first network device does not receive the retrieval terminal device context request message from the second network device, then the first network device performs the reasoning operation.

[0358] In a scenario where a first network device performs an inference operation, the ML model includes a first ML sub-model and a target ML sub-model. Optionally, the ML model also includes a second ML sub-model. On the terminal device side, the model used for inference is described as the "first ML sub-model," and the obtained inference result is described as the "first inference result." On the first network device side, if the first network device performs an inference operation based on the first inference result and obtains a target inference result, the model used for inference on the first network device is described as the "target ML sub-model," and the obtained inference result is described as the "target inference result." For details, see the relevant description of "Second Scenario" below. Alternatively, if the first network device performs an inference operation based on the first inference result and does not obtain the target inference result, the model used for inference on the first network device is described as the "second ML sub-model," and the obtained inference result is described as the "second inference result." For details, see the relevant description of "First Scenario" below. On the second network device side, the model used for inference is described as the "target ML sub-model," and the obtained inference result is described as the "target inference result." Optionally, when inference-related information is transmitted via a CRB, the CRB between the terminal device and the first network device is described as the "first CRB," and the CRB between the terminal device and the second network device is described as the "target CRB."

[0359] Below, taking the switching of the terminal device (i.e., the terminal device switches from the first network device to the second network device, at this time, the first network device is the first access network device, and the second network device is the second access network device) as an example, the third collaborative reasoning method provided in the embodiment of the present application is introduced. This collaborative reasoning method is applied in the machine learning process. Figure 11 The collaborative reasoning method includes S400 to S404, and S800 and the following steps:

[0360] It should be noted that, optionally, in the switching scenario, the process of "configuring the target CRB" (ie, the specific implementation process of S800) can be found in Figure 9a The relevant instructions will not be repeated here.

[0361] S1101: A first network device sends information about a target ML sub-model to a second network device. Correspondingly, the second network device receives the information about the target ML sub-model from the first network device.

[0362] In this case, Figure 11 The target ML sub-model in the scenario is Figure 4 (or Figure 8) is different from the target ML sub-model in the ML model. Here, the ML model includes a first ML sub-model, a second ML sub-model and a target ML sub-model. That is, the output data of the first ML sub-model corresponds to the input data of the second ML sub-model, and the output data of the second ML sub-model corresponds to the input data of the target ML sub-model. In other words, after the first network device divides the ML model to obtain the first ML sub-model, the ML model is further divided to obtain the second ML sub-model and the target ML sub-model. Here, the introduction of the "second ML sub-model" can be referred to the relevant description of S1103a in the first case, which will not be repeated here. For example, still taking Figure 1 Taking the ML model shown as an example, still taking "the first ML sub-model includes the input layer and the first hidden layer" as an example, when the second ML sub-model includes the second hidden layer, the target ML sub-model includes the third hidden layer and the output layer of the ML model.

[0363] For example, the specific implementation process of S1101 can refer to the relevant description of S801, which will not be repeated here.

[0364] It should be noted that S1101 is an optional step. Here, when the first network device performs an inference operation based on the first inference result and fails to obtain the target inference result, the first network device executes S1101. Please refer to the relevant description of "the first case" below for details. Conversely, when the first network device performs an inference operation based on the first inference result and obtains the target inference result, the first network device does not need to execute S1101. Please refer to the relevant description of "the second case" below for details. For the terminal device, the terminal device executes S404 to obtain the first inference result. Afterwards, the terminal device executes S1102:

[0365] S1102: The terminal device sends all information about the first inference result to the first network device. That is, the terminal device sends the complete first inference result to the first network device. Correspondingly, the first network device receives all information about the first inference result from the terminal device.

[0366] The specific implementation process of S1102 can be found in Figure 8 The relevant description of S802a in the second case will not be repeated here.

[0367] It should be noted that the first network device may execute S1101 first and then execute S1102, or may execute S1102 first and then execute S1101, or may execute S1101 and S1102 simultaneously, and this embodiment of the present application is not limited thereto. Furthermore, when the "target ML sub-model" is carried in the handover request message, the first network device executes S1102 first and then executes S1101.

[0368] For the first network device, after obtaining all information about the first inference result, the first network device performs local inference. There are two cases for the local inference performed by the first network device:

[0369] The first case (such as Figure 11 (As shown in the "First Case" block diagram): When the first network device determines that a handover needs to be initiated on the terminal device during the execution of local reasoning, the first network device stops the local reasoning operation process and provides the second reasoning result and the target ML sub-model to the second network device. The second network device then uses the target ML sub-model to continue to perform reasoning operations on the second reasoning result to obtain the target reasoning result. Alternatively, when the first network device determines that a handover needs to be initiated on the terminal device during the execution of local reasoning, and the computing power of the second network device is better than that of the first network device, the first network device still stops the local reasoning operation process and provides the second reasoning result to the second network device. The second network device then continues to perform reasoning operations based on the second reasoning result. In this case, the ML model includes the first ML sub-model, the second ML sub-model and the target ML sub-model. For details, see the relevant descriptions of S1103a to S1103c:

[0370] S1103a. The first network device calculates a second inference result based on all information of the first inference result and the second ML sub-model.

[0371] The input data of the second ML sub-model corresponds to the output data of the first ML sub-model.

[0372] For example, Figure 1 Taking the ML model shown in the figure as an example, when the first ML sub-model includes an input layer and the first hidden layer, the first inference result is the inference result of the first hidden layer. The second ML sub-model includes the second hidden layer. The first network device uses the inference result of the first hidden layer as input data for the second ML sub-model, obtaining the inference result of the second hidden layer, i.e., the second inference result.

[0373] S1103b: The first network device sends the second inference result to the second network device. Correspondingly, the second network device receives the second inference result from the first network device.

[0374] Exemplarily, when the second ML sub-model includes a second hidden layer, the second inference result is an inference result of the second hidden layer. The first network device sends the inference result of the second hidden layer to the second network device.

[0375] S1103c. The second network device calculates a target inference result based on the second inference result and the target ML sub-model.

[0376] The input data of the target ML sub-model corresponds to the output data of the second ML sub-model. The process of the second network device obtaining the target ML sub-model can be referred to the relevant description of S1101 and will not be repeated here.

[0377] For example, Figure 1 Taking the ML model shown in the figure as an example, if the second ML sub-model includes the second hidden layer, the second inference result is the inference result of the second hidden layer. The target ML sub-model includes the third hidden layer and the output layer of the ML model. The second network device uses the inference result of the second hidden layer as input data of the target ML sub-model to obtain the target inference result.

[0378] The second case (such as Figure 11 (As shown in the "Second Case" block diagram): The terminal device switches only after the first network device performs the local inference process. In this way, the first network device performs the local inference operation process and obtains the target inference result. Because the terminal device has switched, the first network device cannot provide the target inference result to the terminal device. Therefore, the first network device provides the target inference result to the second network device, and the second network device provides the target inference result to the terminal device. In this case, the ML model includes the first ML sub-model and the target ML sub-model. For details, see the relevant descriptions of S1103a and S1103b:

[0379] S1103a. The first network device calculates a target inference result based on all information of the first inference result and the target ML sub-model.

[0380] The input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0381] For example, Figure 1 Taking the ML model shown in the figure as an example, when the first ML sub-model includes the first hidden layer, the first inference result is the inference result of the first hidden layer. The target ML sub-model includes the second hidden layer, the third hidden layer, and an output layer. The first network device uses the inference result of the first hidden layer as input data for the target ML sub-model to obtain the target inference result.

[0382] S1103b: The first network device sends the target inference result to the second network device. Correspondingly, the second network device receives the target inference result from the first network device.

[0383] For example, if the target ML sub-model includes a second hidden layer, a third hidden layer, and an output layer, the target inference result is the final inference result of the ML model. The first network device sends the final inference result of the ML model to the second network device. In this case, the first network device provides the target inference result to the second network device. The second network device does not need to obtain the target ML sub-model, that is, the second network device does not need to perform S1101.

[0384] It should be noted that, during the process of executing local reasoning, the first network device determines that the terminal device has switched, and the computing power of the first network device is better than that of the second network device. The first network device can stop the local reasoning operation process and provide the second reasoning result to the second network device, and then the second network device continues to perform the reasoning operation based on the second reasoning result, that is, execute the execution process of the above-mentioned "first case". Alternatively, the first network device can continue to execute the local reasoning operation process, obtain the target reasoning result, and then provide the target reasoning result to the second network device, that is, execute the execution process of the above-mentioned "second case", which is not limited in this embodiment of the present application.

[0385] In the above two cases, the second network device obtains the target inference result in different ways. Then, the second network device executes S1104:

[0386] S1104: The second network device sends the target inference result to the terminal device. Correspondingly, the terminal device receives the target inference result from the second network device.

[0387] The specific implementation process of S1104 can be found in the relevant description of S804 and will not be repeated here.

[0388] It should be noted that in the above steps, if an Xn interface exists between the first network device and the second network device, the relevant information is transmitted between the first and second network devices via the Xn interface. Conversely, if no Xn interface exists between the first and second network devices, the relevant information is transmitted between the first and second network devices via the core network device. The relevant information may include, but is not limited to, the following: information about the target ML submodel, the second inference result, and the target inference result.

[0389] Below, taking the terminal device executing the RRC connection recovery process or the RRC connection re-establishment process as an example, the third collaborative reasoning method provided in the embodiment of the present application is introduced.

[0390] It should be noted that in the "RRC connection recovery" scenario or the "RRC connection re-establishment" scenario, when the first network device is performing local reasoning, if the first network device receives a retrieval terminal device context request message from the second network device, the first network device stops the local reasoning operation process. The first network device provides the second reasoning result to the second network device, and the second network device continues to perform the reasoning operation based on the second reasoning result to obtain the target reasoning result. Alternatively, when the first network device is performing local reasoning, the first network device receives a retrieval terminal device context request message from the second network device, and the computing power of the second network device is better than that of the first network device, the first network device stops the local reasoning operation process, and provides the second reasoning result to the second network device, and the second network device continues to perform the reasoning operation based on the second reasoning result. See details. Figure 11 The specific implementation of the first case.

[0391] Alternatively, after the first network device completes the local reasoning process, if the first network device receives a request message for retrieving the terminal device context from the second network device, the first network device provides the target reasoning result to the second network device. Figure 11 The specific implementation of the second case.

[0392] In addition, in the "RRC connection recovery" scenario or the "RRC connection re-establishment" scenario, when the first network device is performing local reasoning, if the first network device receives a retrieval terminal device context request message from the second network device, and the computing power of the first network device is better than the computing power of the second network device. The first network device can stop the local reasoning operation process, and provide the second reasoning result to the second network device, and then the second network device continues to perform the reasoning operation based on the second reasoning result, that is, execute the execution process of the above-mentioned "first case". Alternatively, the first network device can continue to execute the local reasoning operation process, obtain the target reasoning result, and then provide the target reasoning result to the second network device, that is, execute the execution process of the above-mentioned "second case", and the embodiments of the present application are not limited to this.

[0393] In the third collaborative reasoning method provided in the embodiment of the present application, the terminal device can determine the first reasoning result and send all the information of the first reasoning result to the first network device. The first network device can calculate all the information of the first reasoning result in combination with the target ML sub-model to obtain the target reasoning result, and then provide it to the terminal device through the second network device. Alternatively, the first network device calculates all the information of the first reasoning result in combination with the second ML sub-model to obtain the second reasoning result, and the second network device then calculates the second reasoning result in combination with the target ML sub-model to obtain the target reasoning result, and then provides it to the terminal device. In this way, even if the terminal device switches from the first network device to the second network device, or the terminal device performs RRC connection recovery, or the terminal device performs RRC connection re-establishment, the terminal device does not need to perform a complete reasoning operation, which reduces the delay for the terminal device to obtain the target reasoning result. Similarly, what the terminal device provides to the network device is the intermediate result of the ML model calculation instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0394] The second and third collaborative reasoning methods described above are both described using the scenario of "a first network device determining a first ML sub-model" as an example. Below, still taking the example of a terminal device handover, after the first network device obtains the inference requirement information provided by the terminal device, the first network device determines that the terminal device has switched, and the first network device does not determine the first ML sub-model. Alternatively, after the first network device obtains the inference requirement information provided by the terminal device, the first network device determines that a handover of the terminal device is necessary, and the computing capability of the second network device is superior to that of the first network device, the first network device still does not determine the first ML sub-model, and the second network device determines the first ML sub-model. Taking the example of "RRC connection recovery" or "RRC connection re-establishment" occurring in the terminal device, after the first network device obtains the inference requirement information provided by the terminal device, if the first network device receives a request message from the second network device to retrieve the terminal device context, the first network device does not determine the first ML sub-model. The first network device provides the inference requirement information to the second network device, and the second network device then determines the first ML sub-model. Here, the collaborative reasoning method of the embodiment of the present application is described using the scenario of "the second network device determining the first ML sub-model" as an example. Here, in the scenario where the second network device determines the first ML sub-model, the ML model includes the first ML sub-model and the target ML sub-model. On the terminal device side, the model used to perform reasoning is described as the "first ML sub-model", and the obtained reasoning result is described as the "first reasoning result". On the second network device side, the model used to perform reasoning is described as the "target ML sub-model", and the obtained reasoning result is described as the "target reasoning result". Optionally, in the case of transmitting reasoning-related information through CRB, the CRB between the terminal device and the first network device is described as the "first CRB", and the CRB between the terminal device and the second network device is described as the "target CRB".

[0395] Below, taking the switching of the terminal device (i.e., the terminal device switches from the first network device to the second network device) as an example, the fourth collaborative reasoning method provided by the embodiment of the present application is introduced. This collaborative reasoning method is applied in the machine learning process. Figure 12 The collaborative reasoning method includes S400 and S401, and S800 and the following steps:

[0396] It should be noted that, optionally, when transmitting inference related information through CRB, the process of "configuring the target CRB" can be found in Figure 9a The relevant instructions will not be repeated here.

[0397] S1201: A first network device sends reasoning requirement information to a second network device. Correspondingly, the second network device receives the reasoning requirement information from the first network device.

[0398] For the relevant introduction of “inference requirement information”, please refer to the relevant description of S401, which will not be repeated here.

[0399] Here, in the "handover" scenario, the inference requirement information may be carried in a handover request message. The handover request message is used to request that the terminal device be handed over to the second network device.

[0400] S1202. The second network device determines a first ML sub-model according to the inference requirement information.

[0401] The specific implementation process of S1202 can be found in the relevant description of S402 and will not be repeated here.

[0402] S1203: The second network device sends information about the first ML sub-model to the terminal device via the first network device. Correspondingly, the terminal device receives information about the first ML sub-model from the second network device via the first network device.

[0403] The first ML sub-model is used by the terminal device to perform an inference operation to obtain a first inference result. Here, S1203 is specifically as follows Figure 12 The following describes the specific implementation of S1203 through two possible implementation methods:

[0404] As a first possible implementation, Figure 13 As shown in the block diagram of "a first possible implementation method", when the ML models of the second network device and the terminal device are synchronized, the second network device indicates the first ML sub-model through the first target indication information. Here, "the ML model synchronization between the second network device and the terminal device" means that the meaning represented by the segmentation option of the ML model is applicable to the second network device and the terminal device. In other words, for the second network device and the terminal device, both have the same understanding of the meaning represented by the segmentation option of the ML model. S1203 is specifically implemented as S1203b. About Figure 13 Detailed descriptions of the steps shown are as follows:

[0405] S1203a: The second network device sends model information 1 to the terminal device through the first network device. Correspondingly, the terminal device receives model information 1 from the second network device through the first network device.

[0406] The introduction to model information 1 can be found in the relevant description of S403a and will not be repeated here. Here, the specific implementation process of S1203a is as follows: The second network device sends model information 1 to the first network device. In response, the first network device receives model information 1 from the second network device. Thereafter, the first network device sends model information 1 to the terminal device. In response, the terminal device receives model information 1 from the first network device.

[0407] It should be noted that step S1203a is optional. For example, if the terminal device and the second network device have previously obtained model information 1 from another network device, then step S1203a need not be performed. Alternatively, the terminal device and the second network device can obtain model information 1 from the network control device to achieve model synchronization between the terminal device and the second network device. The network control device may be an OAM device.

[0408] S1203b: The second network device sends the first target indication information to the terminal device through the first network device. Correspondingly, the terminal device receives the first target indication information from the second network device through the first network device.

[0409] The first target indication information can be found in the description of S403b and will not be further described here. The specific implementation process of S1203b is as follows: the second network device sends the first target indication information to the first network device. In response, the first network device receives the first target indication information from the second network device. Thereafter, the first network device sends the first target indication information to the terminal device. In response, the terminal device receives the first target indication information from the first network device.

[0410] S1203c: The terminal device determines a first ML sub-model according to the model information 1 and the first target indication information.

[0411] The specific implementation process of S1203c can be found in the introduction of S403c and will not be repeated here.

[0412] In this way, the second network device sends model information 1 to the terminal device via the first network device to indicate the segmentation positions corresponding to the segmentation options of the ML model, thereby achieving ML model synchronization between the second network device and the terminal device. Furthermore, the second network device can send the first target indication information (i.e., the segmentation options corresponding to the first ML sub-model) to the terminal device via the first network device, allowing the terminal device to determine the first ML sub-model, thereby saving transmission resources.

[0413] As a second possible implementation, Figure 13 As shown in the block diagram of "Second Possible Implementation", when ML model synchronization is not performed, S1203 is specifically implemented as S1203a:

[0414] S1203a: The second network device sends the first ML sub-model full information to the terminal device via the first network device. Correspondingly, the terminal device receives the first ML sub-model full information from the second network device via the first network device.

[0415] Here, the full information of the first ML sub-model refers to information that can fully describe the first ML sub-model, for example, the source code describing the first ML sub-model, the executable program code of the first ML sub-model, or the code of the first ML sub-model that has been partially or fully compiled, etc. In other words, there is no need to synchronize models between the terminal device and the second network device, and the second network device provides the full information of the first ML sub-model to the terminal device through the first network device. Here, the specific implementation process of S1203a is as follows: the second network device sends the full information of the first ML sub-model to the first network device. Accordingly, the first network device receives the full information of the first ML sub-model from the second network device. Afterwards, the first network device sends the full information of the first ML sub-model to the terminal device. Accordingly, the terminal device receives the full information of the first ML sub-model from the first network device.

[0416] S1204. The terminal device calculates a first inference result according to the first ML sub-model.

[0417] The specific implementation process of S1204 can be found in the relevant description of S404 and will not be repeated here.

[0418] S1205: The terminal device sends the first inference result to the second network device. Correspondingly, the second network device receives the first inference result from the terminal device.

[0419] Here, the first reasoning result refers to the complete first reasoning result. Figure 8 The relevant description of S802a in the third case will not be repeated here.

[0420] S1206: The second network device calculates a target inference result based on the first inference result and the target ML sub-model.

[0421] The target ML sub-model includes at least the output layer of the ML model, and the input data of the target ML sub-model corresponds to the output data of the first ML sub-model. Figure 1 Taking "the first ML sub-model includes the input layer and the first hidden layer of the ML model" as an example, the target ML sub-model includes the second hidden layer, the third hidden layer, and the output layer of the ML model.

[0422] Among them, the target inference result is the final inference result of the ML model.

[0423] Exemplarily, the second network device inputs all information about the first inference result into the target ML sub-model, and uses the target ML sub-model to perform processing on the second hidden layer, the third hidden layer, and the output layer to obtain the target inference result. The specific implementation process of S1206 can be found in the relevant description of S803 and will not be repeated here.

[0424] S1207: The second network device sends the target inference result to the terminal device. Correspondingly, the terminal device receives the target inference result from the second network device.

[0425] Here, the specific implementation process of S1207 can refer to the relevant description of S804, which will not be repeated here.

[0426] It should be noted that in the above steps, if an Xn interface exists between the first network device and the second network device, the relevant information is transmitted between the first and second network devices via the Xn interface. Conversely, if no Xn interface exists between the first and second network devices, the relevant information is transmitted between the first and second network devices via the core network device. The relevant information may include, but is not limited to, the following: inference requirement information and information about the first ML sub-model.

[0427] Next, when the terminal device performs the RRC connection recovery process or the RRC connection re-establishment process, the fourth collaborative reasoning method provided in the embodiment of the present application is also applicable. Compared with the fourth collaborative reasoning method in the above-mentioned switching scenario, the differences include the following descriptions:

[0428] First, when transmitting inference-related information through CRB, the configuration process of the target CRB can be found in Figure 10 The steps shown are not repeated here.

[0429] Second, “the second network device provides the terminal device with information about the first ML sub-model” is specifically implemented as follows: Figure 12 S1208 shown in the "RRC connection recovery / RRC connection re-establishment" block diagram:

[0430] S1208: The second network device sends the information of the first ML sub-model to the terminal device. Correspondingly, the terminal device receives the information of the first ML sub-model from the second network device.

[0431] The first ML sub-model is used by the terminal device to perform reasoning operations to obtain a first reasoning result. Figure 6 The relevant instructions are executed by the second network device Figure 6 The relevant processing steps of the first network device in will not be repeated here.

[0432] In the fourth collaborative reasoning method provided in an embodiment of the present application, even if the terminal device switches from the first network device to the second network device, or the terminal device performs RRC connection recovery, or the terminal device performs RRC connection re-establishment, when the first network device sends reasoning requirement information to the second network device, the second network device can determine the first ML sub-model for the terminal device so that the terminal device obtains the first reasoning result. After the terminal device obtains the first reasoning result, it can send all the information of the first reasoning result to the second network device. The second network device can calculate all the information of the first reasoning result in combination with the target ML sub-model to obtain the target reasoning result, and then provide it to the terminal device, so that the terminal device does not need to perform a complete reasoning operation, reducing the delay for the terminal device to obtain the target reasoning result. Similarly, what the terminal device provides to the network device is the intermediate result of the ML model calculation instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0433] The above descriptions of the collaborative reasoning method of the present embodiment are based on the interaction process between a terminal device and a network device. Next, the case where the access network device is implemented as a separate architecture is described.

[0434] In an embodiment of the present application, the terminal device provides reasoning-related information (such as a first reasoning result) to the first DU, and receives a target reasoning result from the first DU. Here, the ML model includes a first ML sub-model and a target ML sub-model. On the terminal device side, the model used to perform reasoning is described as a "first ML sub-model", and the obtained reasoning result is described as a "first reasoning result". On the first DU side, the model used to perform reasoning is described as a "target ML sub-model", and the obtained reasoning result is described as a "target reasoning result". The target reasoning result is the final reasoning result of the ML model. In a scenario where the access network device is implemented as a separate architecture, at least one of the CU, CU-CP and DAM units is described as a "target unit".

[0435] The embodiment of the present application provides a fifth collaborative reasoning method, which is applied in the machine learning process. The specific implementation process can be found in Figure 4 The steps shown in FIG. 1 are that the first DU executes the relevant steps of the first network device. Figure 4 Compared with the first collaborative reasoning method shown, the differences include the following:

[0436] First, in the scenario where the access network device is implemented as a split architecture, the CRB between the terminal device and the target unit is described as the "first CRB". The specific process of "configuring the first CRB" is as follows Figure 14 As shown:

[0437] S1400a. The target unit determines configuration information of the first CRB.

[0438] Among them, the description of "configuration information of the first CRB" can be found in the relevant description of S400a, which will not be repeated here.

[0439] S1400b: The target unit sends the configuration information of the first CRB to the terminal device via the first DU. Correspondingly, the terminal device receives the configuration information of the first CRB from the target unit via the first DU.

[0440] Exemplarily, the target unit sends configuration information of the first CRB to the first DU. Accordingly, the first DU receives the configuration information of the first CRB from the target unit. Thereafter, the first DU sends the configuration information of the first CRB to the terminal device. Accordingly, the terminal device receives the configuration information of the first CRB from the first DU.

[0441] S1400c. The terminal device configures the first CRB according to the configuration information of the first CRB.

[0442] The specific implementation process of S1400c can be found in the relevant description of S400c, which will not be repeated here.

[0443] In this way, when the terminal device obtains the configuration information of the first CRB, the terminal device can configure the first CRB to transmit reasoning-related information through the first CRB.

[0444] Second, during the transmission of inference-related information (such as inference requirement information and all information of the first inference result), if the terminal device sends information to the first DU, the specific implementation process has the following two methods:

[0445] Method 1: The terminal device sends information directly to the first DU.

[0446] Method 2: The terminal device sends information to the first DU through the target unit. In this method, the terminal device sends information to the target unit through an RRC message. Accordingly, the target unit receives the RRC message from the terminal device. The information sent by the terminal device to the first DU is carried in the RRC message. Afterwards, the target unit determines the information carried in the RRC message. The target unit sends the information carried in the RRC message to the first DU. Accordingly, the first DU receives the information from the target unit. Here, taking the terminal device sending inference requirement information to the first DU as an example, the process of "the terminal device sending information to the first DU" is explained: the terminal device sends inference requirement information to the target unit through an RRC message. Accordingly, the target unit receives the RRC message from the terminal device. Afterwards, the target unit determines the inference requirement information carried in the RRC message. The target unit sends inference requirement information to the first DU. Accordingly, the first DU receives the inference requirement information from the target unit.

[0447] Optionally, when the terminal device is configured with the first CRB, the terminal device sends information (such as reasoning requirement information and all information of the first reasoning result) to the target unit via the first CRB. Correspondingly, the target unit receives information from the terminal device via the first CRB.

[0448] If the first DU sends information (such as information about the first ML sub-model and target inference results) to the terminal device, the specific implementation process has the following two methods:

[0449] Method 1: The first DU sends information directly to the terminal device.

[0450] Method 2: The first DU sends information to the terminal device through the target unit. In this case, the first DU sends information to the target unit. Accordingly, the target unit receives the information from the first DU. Afterwards, the target unit sends information to the terminal device through an RRC message. Accordingly, the terminal device receives the RRC message from the target unit. The RRC message carries the information sent by the first DU to the terminal device. Here, taking the first DU sending the target reasoning result to the terminal device as an example, the process of "the first DU sending the target reasoning result to the terminal device" is explained: the first DU sends the target reasoning result to the target unit. Accordingly, the target unit receives the target reasoning result from the first DU. Afterwards, the target unit sends the target reasoning result to the terminal device through an RRC message. Accordingly, the terminal device receives the RRC message from the target unit. The RRC message carries the target reasoning result.

[0451] Optionally, when the terminal device is configured with the first CRB, the target unit sends information (such as information of the first ML sub-model and target inference results) to the terminal device via the first CRB. Correspondingly, the terminal device receives information from the target unit via the first CRB.

[0452] The fifth collaborative reasoning method provided in the embodiment of the present application is that the terminal device uses the first ML sub-model to perform partial reasoning operations to obtain a first reasoning result, and provides the first reasoning result to the first DU. The first DU can operate on all the information of the first reasoning result in combination with the target ML sub-model to obtain the target reasoning result, and then provide it to the terminal device, so that the terminal device does not need to perform a complete reasoning operation, thereby reducing the delay for the terminal device to obtain the target reasoning result. Similarly, the terminal device provides the DU with the intermediate results calculated by the ML model instead of the input data of the ML model, which reduces the risk of "data privacy exposure" and improves the security of the terminal device data.

[0453] In addition, during the transmission of reasoning-related information (such as reasoning requirement information and all information of the first reasoning result), if the terminal device switches, that is, the terminal device switches from the first DU to the second DU, the terminal device receives the target reasoning result from the second DU. In this case, the specific implementation process of the collaborative reasoning method of the embodiment of the present application can be found in Figure 8 、 Figure 11 or Figure 12 The processing steps shown are that the first DU performs the processing steps of the first network device, and the second DU performs the processing steps of the second network device. Figure 12 In the case of the processing steps shown in FIG, “the second DU provides the terminal device with the information of the first ML sub-model” is specifically implemented as follows Figure 12 S1203 shown in the block diagram of “Switching Scenario”, that is, “the second DU provides the information of the first ML sub-model to the terminal device through the first DU”.

[0454] It should be noted that when the first DU sends relevant information (such as information of the target ML sub-model, the first part of the first inference result, all the information of the first inference result, the second inference result, and the target inference result) to the second DU, the specific implementation method can be, for example, but not limited to, the following two methods:

[0455] Method 1: The first DU directly sends relevant information to the second DU. Correspondingly, the second DU directly receives relevant information from the first DU.

[0456] Method 2: The first DU sends relevant information to the second DU via the target unit. Correspondingly, the second DU receives relevant information from the first DU via the target unit.

[0457] Here, when the first DU provides relevant information to the target unit, the target unit sends relevant information to the second DU. Taking the target unit as a CU as an example, if the first DU and the second DU correspond to the same CU, that is, the first DU and the second DU both have interfaces connected to the same CU, then the first DU sends relevant information to the target unit through the F1 interface, and after the target unit receives the relevant information, it sends relevant information to the second DU through the F1 interface. Here, if the first DU and the second DU correspond to different CUs, that is, the first DU corresponds to the first CU and the second DU corresponds to the second CU, then the first DU sends relevant information to the first CU through the F1 interface, the first CU sends relevant information to the second CU through the Xn interface, and the second CU sends relevant information to the second DU through the F1 interface.

[0458] Furthermore, in the case of the second method, Figure 8 or Figure 11 Taking the scenario shown as an example, in the process of the target unit sending the second target indication information to the second DU, the second target indication information can be carried in the terminal device context establishment request message. The terminal device context establishment request message is used to request the second DU to establish the context of the terminal device. Optionally, after the second DU completes the context establishment process, the second DU sends a terminal device context establishment response message to the target unit. Figure 12 Taking the scenario shown as an example, when the target unit sends inference requirement information to the second DU, the inference requirement information can be carried in a terminal device context establishment request message. After the second DU completes the context establishment process, the second DU sends a terminal device context establishment response message to the target unit. Information about the first ML sub-model can also be carried in the terminal device context establishment response message.

[0459] On the other hand, when the second DU sends relevant information (such as model information 1, model information 2, and information of the first ML sub-model) to the first DU, the specific implementation method can be, for example but not limited to, the following two methods: the second DU directly sends the relevant information to the first DU. Alternatively, the second DU sends the relevant information to the first DU through the target unit.

[0460] In the above steps, when the target unit is implemented as a DAM unit, the DAM unit can transmit information with the first DU (or second DU), or transmit information with the first DU (or second DU) through the CU, or transmit information with the first DU (or second DU) through the CU-CP. The target unit and the first DU (or second DU) can use the existing protocol stack to transmit relevant information, or use Figure 15The protocol stack shown transmits relevant information. For example, messages between the target unit and the first DU (or second DU) are carried in high-level data analytics protocol type c (HDAPc) messages. The HDAPc protocol supports functions such as computing data transmission (such as data segmentation and data sorting) between the target unit and the first DU (or second DU), as well as computing data security (such as data integrity protection, data encryption, and data decryption). HDAPc messages can be carried in F1AP messages.

[0461] here, Figure 15 The communication protocol stack between the DU and the target unit is shown. This protocol stack is used to transmit information related to inference operations between the DU and the target unit. The protocol stack mainly includes the HDAPc layer, the F1 application protocol (F1AP) layer, the SCTP layer, the IP layer, the L2 layer, and the L1 layer.

[0462] The above mainly introduces the solution provided by the embodiment of the present application from the perspective of interaction between various network elements. Accordingly, the embodiment of the present application also provides a communication device, which can be the network element in the above method embodiment, or a device including the above network element, or a component that can be used for a network element. It can be understood that in order to implement the above functions, the communication device includes a hardware structure and / or software module corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in hardware or in a computer software-driven hardware manner depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0463] Figure 16 FIG1 shows a schematic structural diagram of a communication device 1600 . The communication device 1600 includes a communication unit 1603 and a processing unit 1602 .

[0464] In the interaction process between the terminal device and the network device, for example, the communication device 1600 is used as the embodiment of the above method. Figure 4 (or Figure 8 、 Figure 11 、 Figure 12), processing unit 1602 is configured to determine a first inference result based on a first machine learning (ML) sub-model. The first ML sub-model is a portion of the ML model. Communication unit 1603 is configured to send the first inference result. Communication unit 1603 is further configured to receive a target inference result. The target inference result is an inference result of the ML model determined based on the first inference result.

[0465] In one possible design, when the communication device 1600 accesses the first network device before determining the first inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the first network device, and receive the target inference result from the first network device, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0466] In one possible design, the communication unit 1603 is further configured to: receive information of the first ML sub-model from the first network device.

[0467] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further used to: receive first model information from the first network device, wherein the first model information includes a correspondence between first candidate indication information and a first segmentation position, and there is at least one first candidate indication information and one first segmentation position; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is a first segmentation position that has a corresponding relationship with one first candidate indication information. The processing unit 1602 is also used to determine the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

[0468] In one possible design, the communication unit 1603 is further used to: send inference requirement information to the first network device, wherein the inference requirement information includes time information when the communication device 1600 obtains the target inference result; and the inference requirement information is used to determine information of the first ML sub-model.

[0469] In one possible design, when communication device 1600 accesses a first network device before sending a first inference result and accesses a second network device while communication device 1600 is sending the first inference result, communication unit 1603 is specifically configured to: send a first portion of information of the first inference result to the first network device, and send a second portion of information of the first inference result to the second network device. Communication unit 1603 is specifically configured to: receive a target inference result from the second network device, where the target inference result is an inference result of an ML model determined based on the first portion of information and the second portion of information.

[0470] In one possible design, when the communication device 1600 accesses the first network device before sending the first inference result, and accesses the second network device after the communication device 1600 sends the first inference result and before receiving the target inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the first network device, and receive the target inference result from the second network device, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0471] In one possible design, when the communication device 1600 accesses the second network device before sending the first inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the second network device, and receive the target inference result from the second network device, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0472] In one possible design, when the communication apparatus 1600 accesses the first network device before determining the first inference result, the communication unit 1603 is further configured to: receive information of the first ML sub-model from the first network device.

[0473] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further used to: receive first model information from the first network device, wherein the first model information includes a correspondence between first candidate indication information and a first segmentation position; there is at least one first candidate indication information and at least one first segmentation position; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is a first segmentation position that has a correspondence with one first candidate indication information. The processing unit 1602 is further used to determine the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

[0474] In one possible design, the communication unit 1603 is further used to: send inference requirement information to the first network device, wherein the inference requirement information includes time information when the communication device 1600 obtains the target inference result; and the inference requirement information is used to determine information of the first ML sub-model.

[0475] In one possible design, when the communication device 1600 accesses the second network device before determining the first inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the second network device, and receive the target inference result from the second network device, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0476] In one possible design, the communication unit 1603 is further configured to: receive information of the first ML sub-model from the first network device, wherein the target network device is the first network device or the second network device.

[0477] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further configured to: receive first model information from the first network device, wherein the first model information includes a correspondence between first candidate indication information and a first segmentation position; there is at least one first candidate indication information and at least one first segmentation position; one first candidate indication information indicates that the ML model is to be segmented, and the position where the ML model is to be segmented is a first segmentation position that has a correspondence with one first candidate indication information; and the processing unit 1602 is further configured to determine the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

[0478] In one possible design, the communication unit 1603 is further configured to: receive information of the first ML sub-model from a second network device, wherein the target network device is the first network device or the second network device.

[0479] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further configured to: receive first model information from the second network device, wherein the first model information includes a correspondence between first candidate indication information and a first segmentation position; there is at least one first candidate indication information and at least one first segmentation position; one first candidate indication information indicates that the ML model is to be segmented, and the position where the ML model is to be segmented is a first segmentation position that has a correspondence with one first candidate indication information; and the processing unit 1602 is further configured to determine the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

[0480] In one possible design, the communication unit 1603 is further used to: send inference requirement information to the first network device, wherein the inference requirement information includes time information when the communication device 1600 obtains the target inference result; and the inference requirement information is used to determine information of the first ML sub-model.

[0481] In the interaction process between the terminal device and the network device, for example, the communication device 1600 is used as the embodiment of the above method. Figure 8 or Figure 11Taking the first network device as an example, the communication unit 1603 is used to receive first reasoning information from the terminal device. The first reasoning information includes all or part of the first reasoning result, the first reasoning result is the reasoning result of the first machine learning ML sub-model, and the first ML sub-model is a part of the ML model. The communication unit 1603 is also used to send second reasoning information to the second network device. The second reasoning information is used to determine the target reasoning result of the ML model, or the second reasoning information is the target reasoning result. The processing unit 1602 is used to determine the second reasoning information based on the first reasoning information.

[0482] In one possible design, the processing unit 1602 is further configured to determine information of the first ML sub-model. The communication unit 1603 is further configured to send the information of the first ML sub-model to the terminal device.

[0483] In one possible design, communication unit 1603 is further configured to receive inference requirement information from a terminal device. The inference requirement information includes an identifier of the ML model and time information when the terminal device obtains a target inference result. Processing unit 1602 is specifically configured to determine information about the first ML sub-model based on the inference requirement information.

[0484] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further configured to send the first model information to the terminal device. The first model information includes a correspondence between the first candidate indication information and the first segmentation position. There is at least one first candidate indication information and one first segmentation position; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is a first segmentation position that corresponds to one first candidate indication information. The first model information and the first target indication information are used by the terminal device to determine the first ML sub-model.

[0485] In one possible design, the first inference information includes all information about the first inference result. Processing unit 1602 is further configured to determine a target inference result based on all information about the first inference result and the second ML sub-model. The second inference information is the target inference result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model.

[0486] In one possible design, the first inference information is identical to the second inference information. Communication unit 1603 is further configured to send information about a target ML sub-model to the second network device. The input data of the target ML sub-model corresponds to the output data of the first ML sub-model. The target ML sub-model is used by the second network device to determine a target inference result.

[0487] In one possible design, the first inference information includes all information about the first inference result. Processing unit 1602 is further configured to determine a second inference result based on all information about the first inference result and the second ML sub-model. The second inference information is the second inference result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model.

[0488] In one possible design, communication unit 1603 is further configured to send information about a target ML sub-model to the second network device, wherein input data of the target ML sub-model corresponds to output data of the second ML sub-model. The target ML sub-model is used by the second network device to determine a target inference result.

[0489] In one possible design, the information of the target ML sub-model includes second target indication information. The communication unit 1603 is further used to receive second model information from the second network device. The second model information includes the correspondence between the second candidate indication information and the second segmentation position. There is at least one second candidate indication information and one second segmentation position; one second candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is a second segmentation position that has a corresponding relationship with one second candidate indication information. The processing unit 1602 is also used to determine the second target indication information from the second candidate indication information based on the correspondence between the second candidate indication information and the second segmentation position, and the target ML sub-model.

[0490] In the interaction process between the terminal device and the network device, for example, the communication device 1600 is used as the embodiment of the above method. Figure 4 The first network device in Figure 8 The second network device in Figure 11 Taking the second network device in the example, communication unit 1603 is used to obtain third inference information. The third inference information is determined based on all information of the first inference result, and the first inference result is an inference result after the operation of the first machine learning ML sub-model, and the first ML sub-model is a part of the ML model. Communication unit 1603 is also used to send a target inference result to the terminal device, where the target inference result is the inference result of the ML model determined based on the third inference information. Processing unit 1602 is used to determine the target inference result based on the third inference information.

[0491] In one possible design, when a terminal device is connected to communication device 1600 before communication device 1600 obtains the third inference information, the third inference information is the complete information of the first inference result; communication unit 1603 is specifically configured to receive the complete information of the first inference result from the terminal device. Processing unit 1602 is further configured to determine a target inference result based on the complete information of the first inference result and a target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0492] In one possible design, the communication unit 1603 is specifically configured to send information of the first ML sub-model to the terminal device.

[0493] In one possible design, the communication unit 1603 is further configured to: receive inference requirement information from the terminal device, wherein the inference requirement information includes time information when the terminal device obtains the target inference result. The processing unit 1602 is further configured to determine information of the first ML sub-model based on the inference requirement information.

[0494] In one possible design, when a terminal device is connected to communication device 1600 during the process of communication device 1600 acquiring the third inference information, the third inference information is the entire information of the first inference result; communication unit 1603 is specifically configured to: receive a first portion of information of the first inference result from the terminal device, and receive a second portion of information of the first inference result from the first network device. Processing unit 1602 is further configured to determine a target inference result based on the first portion of information, the second portion of information, and a target ML sub-model, wherein the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0495] In one possible design, when a terminal device accesses communication device 1600 after communication device 1600 obtains the third inference information, the third inference information is the complete information of the first inference result. Communication unit 1603 is specifically configured to receive the complete information of the first inference result from the first network device. Processing unit 1602 is further configured to determine a target inference result based on the complete information of the first inference result and a target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0496] In one possible design, when the terminal device is connected to communication device 1600 via the first network device before communication device 1600 obtains the third inference information, the third inference information is the complete information of the first inference result. Communication unit 1603 is specifically configured to receive the complete information of the first inference result from the terminal device. Processing unit 1602 is further configured to determine a target inference result based on the complete information of the first inference result and the target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0497] In one possible design, the third inference information is a second inference result, where the second inference result is an inference result of a second ML sub-model determined based on all information from the first inference result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model. Communication unit 1603 is specifically configured to receive the second inference result from the first network device. Processing unit 1602 is further configured to determine a target inference result based on the second inference result and a target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the second ML sub-model.

[0498] In one possible design, after the communication apparatus 1600 obtains the information of the target ML sub-model, when the terminal device accesses the communication apparatus 1600 , the communication unit 1603 is specifically configured to receive the information of the target ML sub-model from the first network device.

[0499] In one possible design, the target ML sub-model information includes second target indication information. Communication unit 1603 is further configured to: send second model information to the first network device, wherein the second model information includes a correspondence between second candidate indication information and a second split position; there is at least one second candidate indication information and at least one second split position; one second candidate indication information indicates that the ML model is to be split, and the position where the ML model is to be split is a second split position that has a correspondence relationship with one second candidate indication information; and the second model information is used by the first network device to determine the second target indication information.

[0500] In one possible design, the third reasoning information is a target reasoning result. The communication unit 1603 is specifically configured to receive the target reasoning result from the first network device.

[0501] In one possible design, during the process of the communication device 1600 sending the information of the first ML sub-model, the communication unit 1603 is specifically used to: send the information of the first ML sub-model to the terminal device; or, send the information of the first ML sub-model to the first network device.

[0502] In one possible design, communication unit 1603 is further configured to: receive inference requirement information from the first network device, wherein the inference requirement information includes time information when the terminal device obtains the target inference result. Processing unit 1602 is further configured to determine information of the first ML sub-model based on the inference requirement information.

[0503] In the scenario where the access network device is implemented as a separate architecture, for example, the communication device 1600 is used as the embodiment of the above method. Figure 4 Taking the terminal device as an example, processing unit 1602 is configured to determine a first inference result based on a first machine learning (ML) sub-model. The first ML sub-model is a portion of the ML model. Communication unit 1603 is configured to send the first inference result. Communication unit 1603 is also configured to receive a target inference result. The target inference result is an inference result of the ML model determined based on the first inference result.

[0504] In one possible design, when the communication device 1600 accesses the first DU before determining the first inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the first DU, and receive the target inference result from the first DU, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0505] In one possible design, the communication unit 1603 is further configured to receive information of the first ML sub-model from the first DU.

[0506] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further used to: receive the first model information from the first DU, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position; the first candidate indication information and the first segmentation position are at least one; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with one first candidate indication information. The processing unit 1602 is further used to determine the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

[0507] In one possible design, the communication unit 1603 is further used to: send reasoning requirement information to the first DU, wherein the reasoning requirement information includes time information when the communication device 1600 obtains the target reasoning result; and the reasoning requirement information is used to determine information of the first ML sub-model.

[0508] In one possible design, when the communication device 1600 accesses the first DU before sending the first inference result and accesses the second DU while the communication device 1600 is sending the first inference result, the communication unit 1603 is specifically configured to: send the first partial information of the first inference result to the first DU, and send the second partial information of the first inference result to the second DU. The communication unit 1603 is specifically configured to: receive a target inference result from the second DU, where the target inference result is an inference result of an ML model determined based on the first partial information and the second partial information.

[0509] In one possible design, when the communication device 1600 accesses the first DU before sending the first inference result, and accesses the second DU after the communication device 1600 sends the first inference result and before receiving the target inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the first DU, and receive the target inference result from the second DU, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0510] In one possible design, when the communication device 1600 accesses the second DU before sending the first inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the second DU, and receive the target inference result from the second DU, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0511] In one possible design, when the communication device 1600 accesses the first DU before determining the first inference result, the communication unit 1603 is further used to: receive information of the first ML sub-model from the first DU.

[0512] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further used to: receive the first model information from the first DU, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position; the first candidate indication information and the first segmentation position are at least one; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with one first candidate indication information. The processing unit 1602 is further used to determine the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

[0513] In one possible design, the communication unit 1603 is further used to: send reasoning requirement information to the first DU, wherein the reasoning requirement information includes time information when the communication device 1600 obtains the target reasoning result; and the reasoning requirement information is used to determine information of the first ML sub-model.

[0514] In one possible design, when the communication device 1600 accesses the second DU before determining the first inference result, the communication unit 1603 is specifically used to: send all information of the first inference result to the second DU, and receive the target inference result from the second DU, wherein the target inference result is the inference result of the ML model determined based on all information of the first inference result.

[0515] In one possible design, during the process of the communication device 1600 acquiring the information of the first ML sub-model, the communication unit 1603 is specifically configured to: receive the information of the first ML sub-model from the first DU.

[0516] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further used to: receive the first model information from the first DU, wherein the first model information includes the correspondence between the first candidate indication information and the first segmentation position; the first candidate indication information and the first segmentation position are at least one; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a corresponding relationship with one first candidate indication information. The processing unit 1602 is further used to determine the first ML sub-model based on the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

[0517] In one possible design, the communication unit 1603 is further used to: send reasoning requirement information to the first DU, wherein the reasoning requirement information includes time information when the communication device 1600 obtains the target reasoning result; and the reasoning requirement information is used to determine information of the first ML sub-model.

[0518] In the scenario where the access network equipment is implemented as a split architecture, for example, in the first DU Figure 8 or Figure 11In the case of the step of the first network device in the above method embodiment, taking the communication device 1600 as the first DU in the above method embodiment as an example, the communication unit 1603 is used to receive the first reasoning information from the terminal device. The first reasoning information includes all or part of the information of the first reasoning result, the first reasoning result is the reasoning result of the first machine learning ML sub-model, and the first ML sub-model is a part of the ML model. The communication unit 1603 is also used to send the second reasoning information to the second DU. The second reasoning information is used to determine the target reasoning result of the ML model, or the second reasoning information is the target reasoning result. The processing unit 1602 is used to determine the second reasoning information based on the first reasoning information.

[0519] In one possible design, the processing unit 1602 is further configured to determine information of the first ML sub-model. The communication unit 1603 is further configured to send the information of the first ML sub-model to the terminal device.

[0520] In one possible design, communication unit 1603 is further configured to receive inference requirement information from a terminal device. The inference requirement information includes an identifier of the ML model and time information when the terminal device obtains a target inference result. Processing unit 1602 is specifically configured to determine information about the first ML sub-model based on the inference requirement information.

[0521] In one possible design, the information of the first ML sub-model includes first target indication information. The communication unit 1603 is further configured to send the first model information to the terminal device. The first model information includes a correspondence between the first candidate indication information and the first segmentation position. There is at least one first candidate indication information and one first segmentation position; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that corresponds to one first candidate indication information; the first model information and the first target indication information are used by the terminal device to determine the first ML sub-model.

[0522] In one possible design, the first inference information includes all information about the first inference result. Processing unit 1602 is further configured to determine a target inference result based on all information about the first inference result and the second ML sub-model. The second inference information is the target inference result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model.

[0523] In one possible design, the first reasoning information is identical to the second reasoning information. Communication unit 1603 is further configured to send information about a target ML sub-model to the second DU. The input data of the target ML sub-model corresponds to the output data of the first ML sub-model. The target ML sub-model is used by the second DU to determine a target reasoning result.

[0524] In one possible design, the first inference information includes all information about the first inference result. Processing unit 1602 is further configured to determine a second inference result based on all information about the first inference result and the second ML sub-model. The second inference information is the second inference result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model.

[0525] In one possible design, communication unit 1603 is further configured to send information about a target ML sub-model to the second DU, where the input data of the target ML sub-model corresponds to the output data of the second ML sub-model. The target ML sub-model is used by the second DU to determine a target inference result.

[0526] In one possible design, the information of the target ML sub-model includes second target indication information. The communication unit 1603 is further used to receive the second model information from the second DU. The second model information includes the correspondence between the second candidate indication information and the second segmentation position. There is at least one second candidate indication information and one second segmentation position; one second candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is a second segmentation position that has a corresponding relationship with one second candidate indication information. The processing unit 1602 is also used to determine the second target indication information from the second candidate indication information based on the correspondence between the second candidate indication information and the second segmentation position, and the target ML sub-model.

[0527] For example, in the second DU, Figure 4 The first network device, or Figure 8 The second network device, or Figure 11 In the case of the step of the second network device in the above method embodiment, taking the communication device 1600 as the second DU in the above method embodiment as an example, the communication unit 1603 is used to obtain the third reasoning information. The third reasoning information is determined based on all the information of the first reasoning result, and the first reasoning result is the reasoning result after the operation of the first machine learning ML sub-model, and the first ML sub-model is a part of the ML model. The communication unit 1603 is also used to send the target reasoning result to the terminal device, wherein the target reasoning result is the reasoning result of the ML model determined based on the third reasoning information. The processing unit 1602 is used to determine the target reasoning result based on the third reasoning information.

[0528] In one possible design, when a terminal device is connected to communication device 1600 before communication device 1600 obtains the third inference information, the third inference information is the complete information of the first inference result; communication unit 1603 is specifically configured to receive the complete information of the first inference result from the terminal device. Processing unit 1602 is further configured to determine a target inference result based on the complete information of the first inference result and a target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0529] In one possible design, the communication unit 1603 is specifically configured to send information of the first ML sub-model to the terminal device.

[0530] In one possible design, the communication unit 1603 is further configured to: receive inference requirement information from the terminal device, wherein the inference requirement information includes time information when the terminal device obtains the target inference result. The processing unit 1602 is further configured to determine information of the first ML sub-model based on the inference requirement information.

[0531] In one possible design, when a terminal device is connected to communication device 1600 during the process of communication device 1600 acquiring the third inference information, the third inference information is the entire information of the first inference result; and communication unit 1603 is specifically configured to: receive a first portion of information of the first inference result from the terminal device, and receive a second portion of information of the first inference result from the first DU. Processing unit 1602 is further configured to determine a target inference result based on the first portion of information, the second portion of information, and a target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0532] In one possible design, when a terminal device accesses communication device 1600 after communication device 1600 obtains the third inference information, the third inference information is the complete information of the first inference result. Communication unit 1603 is specifically configured to receive the complete information of the first inference result from the first DU. Processing unit 1602 is further configured to determine a target inference result based on the complete information of the first inference result and the target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0533] In one possible design, when the terminal device is connected to communication device 1600 via the first network device before communication device 1600 obtains the third inference information, the third inference information is the complete information of the first inference result. Communication unit 1603 is specifically configured to receive the complete information of the first inference result from the terminal device. Processing unit 1602 is further configured to determine a target inference result based on the complete information of the first inference result and the target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

[0534] In one possible design, the third inference information is a second inference result, where the second inference result is an inference result of a second ML sub-model determined based on all information from the first inference result, and the input data of the second ML sub-model corresponds to the output data of the first ML sub-model. Communication unit 1603 is specifically configured to receive the second inference result from the first DU. Processing unit 1602 is further configured to determine a target inference result based on the second inference result and a target ML sub-model, where the input data of the target ML sub-model corresponds to the output data of the second ML sub-model.

[0535] In one possible design, after the communication apparatus 1600 obtains the information of the target ML sub-model, when the terminal device accesses the communication apparatus 1600 , the communication unit 1603 is specifically configured to receive the information of the target ML sub-model from the first DU.

[0536] In one possible design, the target ML sub-model information includes second target indication information. The communication unit 1603 is further configured to: send second model information to the first DU, wherein the second model information includes a correspondence between second candidate indication information and a second segmentation position; there is at least one second candidate indication information and at least one second segmentation position; one second candidate indication information indicates that the ML model is to be segmented, and the position where the ML model is segmented is a second segmentation position that has a correspondence relationship with one second candidate indication information; and the second model information is used by the first DU to determine the second target indication information.

[0537] In one possible design, the third reasoning information is a target reasoning result. The communication unit 1603 is specifically configured to receive the target reasoning result from the first DU.

[0538] In one possible design, during the process of the communication device 1600 sending the information of the first ML sub-model, the communication unit 1603 is specifically used to: send the information of the first ML sub-model to the first DU.

[0539] In one possible design, the communication unit 1603 is further configured to: receive inference requirement information from the first DU, wherein the inference requirement information includes time information for the terminal device to obtain the target inference result. The processing unit 1602 is further configured to determine information of the first ML sub-model based on the inference requirement information.

[0540] Among them, all relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module and will not be repeated here.

[0541] It should be understood that the processing unit 1602 in the embodiment of the present application can be implemented by a processor or a processor-related circuit component, and the communication unit 1603 can be implemented by a transceiver or a transceiver-related circuit component.

[0542] As a possible implementation, an embodiment of the present application provides a chip comprising a logic circuit and an input / output interface. The input / output interface is used to communicate with modules outside the chip, and the logic circuit is used to perform other operations on the terminal device in addition to the sending and receiving operations in the above method embodiment.

[0543] For example, the chip is implemented as the embodiment of the above method Figure 4 Taking the functions of the terminal device as an example, the input / output interface is used to output the information in S401 and S405 on the terminal device side, and the input / output interface is also used to input the information in S403 and S407 on the terminal device side, and / or the input / output interface is also used to perform other sending and receiving steps on the terminal device side in the embodiment of the present application. The logic circuit is used to perform S404 on the terminal device side, and / or the logic circuit is also used to perform other processing steps on the terminal device side in the embodiment of the present application.

[0544] For example, the chip is implemented as the above method embodiment Figure 8 Taking the functions of the terminal device as an example, the input / output interface is used to output the information in S802a and S802c on the terminal device side, the input / output interface is also used to input the information in S804 on the terminal device side, and / or the input / output interface is also used to perform other sending and receiving steps on the terminal device side in the embodiment of the present application. The logic circuit is used to perform other processing steps on the terminal device side.

[0545] For example, the above method embodiment is implemented as a chip Figure 11 Taking the functions of the terminal device as an example, the input / output interface is used to output the information in S1102 on the terminal device side, the input / output interface is also used to input the information in S1104 on the terminal device side, and / or the input / output interface is also used to perform other sending and receiving steps on the terminal device side in the embodiment of the present application. The logic circuit is used to perform other processing steps on the terminal device side in the embodiment of the present application.

[0546] For example, the above method embodiment is implemented as a chip Figure 12 Taking the functions of the terminal device as an example, the input and output interface is used to input the information in S1203, S1207, and S1208 on the terminal device side, and the input and output interface is also used to output the information in S1205 on the terminal device side, and / or the input and output interface is also used to perform other sending and receiving steps on the terminal device side in the embodiment of the present application. The logic circuit is used to execute S1204 on the terminal device side, and / or the logic circuit is also used to perform other processing steps on the terminal device side in the embodiment of the present application.

[0547] For example, the chip is implemented as the embodiment of the above method Figure 4 Taking the function of the first network device as an example, the input / output interface is used to input the information in S401 and S405 on the first network device side, and the input / output interface is also used to output the information in S403 and S407 on the first network device side, and / or the input / output interface is also used to perform other sending and receiving steps on the first network device side in the embodiment of the present application. The logic circuit is used to perform S402 and S406 on the first network device side, and / or the logic circuit is also used to perform other processing steps on the first network device side in the embodiment of the present application.

[0548] For example, the chip is implemented as the above method embodiment Figure 8 Taking the functions of the first network device as an example, the input / output interface is used to input the information in S802a on the first network device side, and the input / output interface is also used to output the information in S801 and S802b on the first network device side, and / or the input / output interface is also used to perform other sending and receiving steps on the first network device side in the embodiment of the present application. The logic circuit is used to perform other processing steps on the first network device side.

[0549] For example, the above method embodiment is implemented as a chip Figure 11 Taking the functions of the first network device as an example, the input / output interface is used to input the information in S1102 on the first network device side, and the input / output interface is also used to output the information in S1101 and S1103b on the first network device side, and / or the input / output interface is also used to perform other sending and receiving steps on the first network device side in the embodiment of the present application. The logic circuit is used to execute S1103a on the first network device side, and / or the logic circuit is also used to perform other processing steps on the first network device side in the embodiment of the present application.

[0550] For example, the above method embodiment is implemented as a chip Figure 12 Taking the function of the first network device as an example, the input / output interface is used to input the information in S1203 on the first network device side, the input / output interface is used to output the information in S1201 and S1203 on the first network device side, and / or the input / output interface is also used to perform other sending and receiving steps on the first network device side in the embodiment of the present application. The logic circuit is used to perform other processing steps on the first network device side in the embodiment of the present application.

[0551] For example, the above method embodiment is implemented as a chip Figure 8Taking the function of the second network device as an example, the input / output interface is used to input the information in S801, S802a, and S802b on the second network device side, and the input / output interface is also used to output the information in S804 on the second network device side, and / or the input / output interface is also used to perform other sending and receiving steps on the second network device side in the embodiment of the present application. The logic circuit is used to execute S803 on the second network device side, and / or the logic circuit is also used to perform other processing steps.

[0552] For example, the above method embodiment is implemented as a chip Figure 11 Taking the function of the second network device as an example, the input / output interface is used to input the information in S1101 and S1103b on the second network device side, and the input / output interface is also used to output the information in S1104 on the second network device side, and / or the input / output interface is also used to perform other transceiver steps on the second network device side in the embodiment of the present application. The logic circuit is used to execute S1103c on the second network device side, and / or the logic circuit is also used to perform other processing steps on the second network device side in the embodiment of the present application.

[0553] For example, the above method embodiment is implemented as a chip Figure 12 Taking the function of the second network device as an example, the input and output interface is used to input the information in S1201 and S1205 on the second network device side, and the input and output interface is also used to output the information in S1203, S1207, and S1208 on the second network device side, and / or the input and output interface is also used to perform other sending and receiving steps on the second network device side in the embodiment of the present application. The logic circuit is used to execute S1202 and S1206 on the second network device side in the embodiment of the present application, and / or the logic circuit is also used to perform other processing steps on the second network device side in the embodiment of the present application.

[0554] Optionally, the communication device 1600 may further include a storage unit 1601 for storing program codes and data of the communication device 1600 . The data may include but is not limited to original data or intermediate data.

[0555] The processing unit 1602 may be a processor or controller, such as a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. A processor may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0556] The communication unit 1603 can be a communication interface, a transceiver or a transceiver circuit, etc., where the communication interface is a general term. In a specific implementation, the communication interface can include multiple interfaces, for example, it can include: an interface between the first access network device and the second access network device and / or other interfaces.

[0557] The storage unit 1601 may be a memory.

[0558] When the processing unit 1602 is a processor, the communication unit 1603 is a communication interface, and the storage unit 1601 is a memory, the communication device 1700 involved in the embodiment of the present application can be Figure 17 shown.

[0559] See Figure 17 As shown, the communication device 1700 includes: a processor 1702 , a transceiver 1703 , and a memory 1701 .

[0560] The transceiver 1703 may be a standalone transmitter that can be used to send information to other devices, or a standalone receiver that can be used to receive information from other devices. The transceiver may also be a component that integrates the functions of sending and receiving information. The embodiments of the present application do not limit the specific implementation of the transceiver.

[0561] Optionally, the communication device 1700 may further include a bus 1704. The transceiver 1703, the processor 1702, and the memory 1701 may be interconnected via the bus 1704; the bus 1704 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus 1704 may be divided into an addr...

Claims

1. A collaborative reasoning method, characterized in that: include: The terminal device receives first target indication information, where the first target indication information indicates a segmentation position of the machine learning ML model; The terminal device determines a first ML sub-model according to the first target indication information, where the first ML sub-model is a part of the ML model; The terminal device determines a first inference result based on the first ML sub-model; The terminal device sends a first part of information of the first inference result to the first network device; The terminal device switches from the first network device to the second network device; The terminal device sends the second part of information of the first inference result to the second network device; The terminal device receives a target inference result, wherein the target inference result is an inference result of the ML model determined based on the first inference result.

2. The method according to claim 1, characterized in that In the case where the terminal device accesses the first network device before sending the first inference result, and accesses the second network device during the process of sending the first inference result, The terminal device receives the target inference result, including: The terminal device receives the target inference result from the second network device, wherein the target inference result is an inference result of the ML model determined based on the first part of information and the second part of information.

3. The method according to claim 2, characterized in that In a case where the terminal device accesses the first network device before determining the first inference result, the method further includes: The terminal device receives information of the first ML sub-model from the first network device.

4. The method according to claim 3, characterized in that The method further comprises: The terminal device receives first model information from the first network device, wherein the first model information includes a correspondence between first candidate indication information and a first segmentation position; the first candidate indication information and the first segmentation position are both at least one; one first candidate indication information indicates that the ML model is to be segmented, and the position for segmenting the ML model is the first segmentation position that has a correspondence relationship with the one first candidate indication information; The terminal device determines the first ML sub-model according to the first target indication information, including: the terminal device determines the first ML sub-model according to the correspondence between the first candidate indication information and the first segmentation position, and the first target indication information.

5. The method according to claim 3 or 4, characterized in that The method further comprises: The terminal device sends inference requirement information to the first network device, wherein the inference requirement information includes time information when the terminal device obtains the target inference result; the inference requirement information is used to determine information of the first ML sub-model.

6. The method according to any one of claims 1 to 5, characterized in that The input data of the first ML sub-model is data generated by the terminal device.

7. A collaborative reasoning method, characterized in that: include: A first network device sends first target indication information to a terminal device, where the first target indication information indicates a split position of a machine learning (ML) model, and the first target indication information is used to determine a first ML sub-model, where the first ML sub-model is a part of the ML model. The first network device is the network device before the terminal device is switched; The first network device receives a first portion of information of a first inference result from the terminal device, where the first portion of information of the first inference result is an inference result of the first ML sub-model; The first network device sends a first portion of information of the first inference result to a second network device, where the second network device is the network device after the terminal device has switched, and the switching occurs after the terminal device sends the first portion of information of the first inference result to the first network device; The first network device sends second inference information to the second network device, wherein the second inference information is determined based on the first part of information, and the second inference information is used to determine a target inference result of the ML model, or the second inference information is the target inference result.

8. The method according to claim 7, characterized in that The method further comprises: The first network device determines information of the first ML sub-model; The first network device sends information of the first ML sub-model to the terminal device.

9. The method according to claim 8, characterized in that The method further comprises: The first network device receives inference requirement information from the terminal device, wherein the inference requirement information includes time information when the terminal device obtains the target inference result; The first network device determines information about the first ML sub-model, including: The first network device determines information of the first ML sub-model according to the inference requirement information.

10. The method according to claim 8 or 9, characterized in that The method further comprises: The first network device sends first model information to the terminal device, wherein the first model information includes a correspondence between first candidate indication information and a first segmentation position; the first candidate indication information and the first segmentation position are both at least one; one first candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the first segmentation position that has a correspondence with the one first candidate indication information; the first model information is used by the terminal device to determine the first ML sub-model.

11. The method according to any one of claims 7 to 10, characterized in that The method further comprises: The first network device determines the target inference result based on all information of the first inference result and the target ML sub-model, wherein the second inference information is the target inference result, and the input data of the target ML sub-model corresponds to the output data of the first ML sub-model.

12. The method according to any one of claims 7 to 11, characterized in that The method further comprises: The first network device determines a second inference result based on all information of the first inference result and a second ML sub-model, wherein the second inference information is the second inference result, and input data of the second ML sub-model corresponds to output data of the first ML sub-model.

13. The method according to claim 12, characterized in that The method further comprises: The first network device sends information of the target ML sub-model to the second network device; The input data of the target ML sub-model corresponds to the output data of the second ML sub-model; and the target ML sub-model is used by the second network device to determine the target inference result.

14. The method according to any one of claims 7 to 10, characterized in that The method further comprises: The first network device sends information of the target ML sub-model to the second network device; The input data of the target ML sub-model corresponds to the output data of the first ML sub-model; and the target ML sub-model is used by the second network device to determine the target inference result.

15. The method according to claim 13 or 14, characterized in that The information of the target ML sub-model includes second target indication information; the method further includes: The first network device receives second model information from the second network device, wherein the second model information includes a correspondence between second candidate indication information and a second segmentation position; the second candidate indication information and the second segmentation position are both at least one; one second candidate indication information indicates that the ML model is to be segmented, and the position for segmenting the ML model is the second segmentation position that has a correspondence relationship with the one second candidate indication information; The first network device determines the second target indication information from the second candidate indication information according to the correspondence between the second candidate indication information and the second segmentation position, and the target ML sub-model.

16. A collaborative reasoning method, characterized in that include: The second network device receives second target indication information, where the second target indication information indicates a split position of the machine learning ML model, and the second network device is the network device after the terminal device switches; The second network device determines a target ML sub-model according to the second target indication information, where the target ML sub-model is a part of the ML model; The second network device receives a first portion of information of a first inference result from a first network device, where the first network device is a network device before the terminal device switches, and the switching occurs after the terminal device sends the first portion of information of the first inference result to the first network device; The second network device receives second part information of the first inference result from the terminal device, where the first part information and the second part information of the first inference result are inference results of the first ML sub-model; The second network device obtains third reasoning information, where the third reasoning information is determined based on the first part of information and the second part of information of the first reasoning result, where the first reasoning result is a reasoning result after computing a first ML sub-model, where the first ML sub-model is a part of the ML model; The second network device determines a target inference result based on the third inference information and the target ML sub-model; The second network device sends the target inference result to the terminal device, wherein the target inference result is an inference result of the ML model determined based on the third inference information.

17. The method according to claim 16, characterized in that When the terminal device is connected to the second network device during the process of the second network device acquiring the third inference information, The second network device determines a target inference result based on the third inference information and the target ML sub-model, including: The second network device determines the target inference result based on the first portion of information, the second portion of information, and the target ML sub-model, wherein input data of the target ML sub-model corresponds to output data of the first ML sub-model.

18. The method according to claim 16, characterized in that The third inference information is a second inference result, where the second inference result is an inference result of a second ML sub-model determined based on all information of the first inference result, and input data of the second ML sub-model corresponds to output data of the first ML sub-model; The second network device obtains the third inference information, including: The second network device receives the second inference result from the first network device; The second network device determines a target inference result based on the third inference information and the target ML sub-model, including: The second network device determines the target inference result based on the second inference result and the target ML sub-model, wherein input data of the target ML sub-model corresponds to output data of the second ML sub-model.

19. The method according to any one of claims 16 to 18, characterized in that After the second network device acquires the information of the target ML sub-model, when the terminal device is connected to the second network device, the second network device acquires the information of the target ML sub-model, including: The second network device receives information of the target ML sub-model from the first network device.

20. The method according to claim 19, characterized in that The method further comprises: The second network device sends second model information to the first network device, wherein the second model information includes a correspondence between second candidate indication information and a second segmentation position; the second candidate indication information and the second segmentation position are at least one; one second candidate indication information indicates that the ML model is segmented, and the position where the ML model is segmented is the second segmentation position that has a corresponding relationship with the one second candidate indication information; the second model information is used by the first network device to determine the second target indication information.

21. A communication device, characterized in that: include: A unit for performing the steps of any one of claims 1 to 6; Alternatively, the communication device comprises: a unit for performing each step according to any one of claims 7 to 15; Alternatively, the communication device includes: a unit for executing each step described in any one of claims 16 to 20.

22. A communication device, characterized in that: include: A processor and a memory, wherein the processor and the memory are coupled, and the memory stores program instructions. When the program instructions stored in the memory are executed by the processor, the collaborative reasoning method according to any one of claims 1 to 6 is implemented, or the collaborative reasoning method according to any one of claims 7 to 15 is implemented, or the collaborative reasoning method according to any one of claims 16 to 20 is implemented.

23. A chip, characterized in that: The chip includes a logic circuit and an input / output interface, the input / output interface is used to communicate with a module outside the chip, and the logic circuit is used to run a computer program or instruction to implement the collaborative reasoning method as described in any one of claims 1 to 6, or to implement the collaborative reasoning method as described in any one of claims 7 to 15, or to implement the collaborative reasoning method as described in any one of claims 16 to 20.

24. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, and when the program is called by the processor, the collaborative reasoning method according to any one of claims 1 to 6 is executed. or the collaborative reasoning method according to any one of claims 7 to 15 is executed; Or the collaborative reasoning method described in any one of claims 16 to 20 is executed.

Images