Method, apparatus and system for model training

Abstract

Provided in the embodiments of the present disclosure are a method, apparatus and system for model training. The method may comprise: a network management device acquiring training data for training a first part of a bilateral model, wherein the bilateral model comprises the first part to be deployed at a network device and a second part to be deployed at a terminal device; performing model training on the basis of the training data, so as to obtain the first part; and sending the first part to the network device. Exemparily, the training of the second part may be completed by a server or the network management device. In this way, the network management device on a management plane assists in the training process of the bilateral model in an air interface. Specifically, at least the first part is trained and deployed at the network device, so as to support the use of the bilateral model in the air interface.

Description

Method, device, and system for model training

[0001] This application claims priority to the Chinese patent application filed with the Intellectual Property Office of the People's Republic of China on October 23, 2023, with application number 202311376105.2 and invention name "Method, device, and system for model training", the entire contents of which are incorporated herein by reference. Technical Field

[0002] Embodiments of the present disclosure relate to the field of communications, and more particularly, to a method, apparatus, system, and computer-readable storage medium for model training. Background Art

[0003] To enhance network intelligence and automation, artificial intelligence (AI) and machine learning (ML) technologies are being applied in a growing number of fields. The 3rd Generation Partnership Project (3GPP) working group is currently researching several related standardization topics related to network intelligence. For example, the 3GPP Radio Access Network (RAN) working group is currently discussing AI / ML algorithm research for the new radio (NR) air interface. This research focuses on enhancing the air interface with AI / ML-based algorithms to improve air interface performance. There are two ways to use AI / ML on the air interface: unilateral AI and bilateral AI. In this scenario, a model is deployed on one side of the access network equipment or terminal device, and model inference is performed by the party that deployed the model. This AI model usage model is called unilateral AI, and the corresponding model used is called a unilateral AI model. In this scenario, a model is divided into two parts, deployed on one side of the access network equipment or terminal device, and both parties use their own deployed models for inference, collaborating to complete the inference task. This AI model usage model is called bilateral AI, and the corresponding model used is called a bilateral AI model.

[0004] The bilateral AI model includes a part deployed at the access network equipment (for example, called a decoder model) and a part deployed at the terminal device (for example, called an encoder model). However, how to obtain the bilateral AI model through training still requires further research.

[0005] Summary of the Invention

[0006] The embodiments of the present disclosure provide a solution for model training, and specifically provide a training solution for a bilateral AI model.

[0007] In a first aspect of the present disclosure, a method for model training is provided. The method may include: obtaining training data for training a first part of a bilateral model, wherein the bilateral model includes a first part to be deployed at a network device and a second part to be deployed at a terminal device; performing model training based on the training data to obtain the first part; and sending the first part to the network device.

[0008] The embodiments of the present disclosure can be executed by a network management device. In this way, the network management device on the management plane can assist in the training process of the bilateral model of the air interface, specifically training the first part and deploying it on the network device, thereby supporting the use of the bilateral model in the air interface.

[0009] In some implementations, the method may further include performing lightweight processing on the first portion, such that the lightweighted first portion meets at least one of the following: a performance requirement for the first portion, an energy requirement for training the first portion, an energy requirement for an inference task for which the first portion will be used, or a first model size requirement for the first portion. In this way, lightweight processing can be performed while ensuring the requirements of the model to compress the model, thereby reducing the computational complexity and size of the model.

[0010] Exemplarily, the method may be executed by a second network management device, such as an element management system (EMS).

[0011] In some implementations, obtaining training data for training the first part of the bilateral model includes receiving a first training request from a first network management device, the first training request instructing to train the first part of the bilateral model, the first training request including the training data for training the first part of the bilateral model and structural information of the first part. In this manner, based on the first training request from the first network management device, model training can be performed using the training data and structural information provided by the first network management device, thereby obtaining the first part that matches the second part.

[0012] In some implementations, the method may further include: sending a first model training report to the first network management device, the first model training report including at least one of the following: a model identifier of the bilateral model, a model identifier of the first part, actual performance of the first part, actual energy consumption of training the first part, estimated energy consumption of an inference task for which the first part will be used, or an actual model size of the first part. In this manner, by sending the first model training report to the first network management device, the first network management device can determine that training of the first part has been completed, and can also obtain information about the training status of the first part, such as actual energy consumption, actual model size, etc.

[0013] In some implementations, obtaining training data for training the first part of the bilateral model includes obtaining training data for training the first part from a network device. Exemplarily, the method further includes receiving a first training request from a first network management device, the first training request requesting training of the first part of the bilateral model and also requesting training data for training the second part. Exemplarily, performing model training based on the training data to obtain the first part includes: obtaining the first part through training based on the first training request using the training data for training the first part; determining structural information of the second part, and constructing training data for training the second part. Exemplarily, the method may further include sending a first model training report to the first network management device, the first model training report including the training data for training the second part and the structural information of the second part. In this manner, the training data for training the second part and the structural information of the second part can be subsequently provided to a server, which trains the second part.

[0014] In some implementations, obtaining training data for training the first part of the bilateral model includes obtaining training data for training the first part and the second part from a network device. For example, during model training, the first part and the second part can be obtained based on the training data. In this way, both parts of the bilateral model can be trained by the network management device, reducing signaling interactions.

[0015] In some implementations, the method may further include: receiving a model loading request from a first network management device, the model loading request instructing that the first portion be loaded onto the network device; and sending the first portion to the network device based on the model loading request. In this manner, the first portion may be loaded onto the network device based on the model loading request from the first network management device, thereby ensuring that the network device treats the first portion as part of a bilateral model rather than a unilateral model.

[0016] In some implementations, the model loading request may include at least one of the following: a first identifier of the network device, a model identifier of the first part, a second identifier of the terminal device, a model identifier of the second part, or a model identifier of the bilateral model.

[0017] In some implementations, the first training request also includes at least one of the following: a model identifier of the bilateral model, a model identifier of the first part, a model identifier of the second part, a first identifier of the network device, a second identifier of the terminal device, indication information for indicating the type of reasoning task for which the bilateral model will be used, performance requirements of the bilateral model, energy consumption requirements for bilateral model training, energy consumption requirements for the reasoning task, a first model size requirement for the first part, or a lightweight indication.

[0018] In this way, by including the model identifiers of the first and / or second parts in the request, the matching and association between the first and second parts can be determined. By including the identifiers of the network device and / or terminal device in the request, the two parties to deploy the bilateral model can be determined. By including the requirements for the model in the first training request, the actual available resources can be considered during the training process, ensuring the availability of the trained model.

[0019] In some implementations, the method may further include: receiving a request message from a first network management device, the request message being for status information of the network device; and, in response to the request message, sending a response message to the first network management device, the response message including at least one of the following: computing power status of the network device, storage status of the network device, energy consumption information of the network device, or a network energy consumption target value of the network device. In this manner, the first network management device can determine, for example, training requirements based on the status information, thereby ensuring the availability of the trained model.

[0020] In some implementations, the method may further include: sending a second model training report to the first network management device, where the second model training report includes training data used for training the second portion and structural information about the second portion. In this manner, the training data used for training the second portion may be further provided to the server, so that the server can train the second portion to obtain a matching portion.

[0021] Exemplarily, the method may be executed by a first network management device, such as a network management system (NMS).

[0022] In some implementations, obtaining training data for training the first part of the bilateral model includes receiving a second training request from a server, the second training request instructing to train the first part of the bilateral model, the second training request including the training data for training the first part of the bilateral model and structural information of the first part. In this manner, based on the second training request from the server, model training can be performed using the training data and structural information provided by the server, thereby obtaining the first part that matches the second part.

[0023] In some implementations, the method may further include sending a fourth model training report to the server, where the fourth model training report includes at least one of the following: a model identifier of the bilateral model, a model identifier of the first part, or actual performance of the first part. In this manner, sending the fourth model training report to the server enables the server to determine that training of the first part has been completed, thereby enabling the server to further decide to deploy the model.

[0024] In some implementations, the method may further include: receiving a notification message from the server indicating that the second part trained by the server has been loaded onto the terminal device; and, in response to the notification message, sending the first part to the network device. In this manner, the first part may be loaded onto the network device based on the notification message from the server after the second part has been deployed onto the terminal device, thereby ensuring that the network device utilizes the first part as part of a bilateral model rather than a unilateral model.

[0025] In some implementations, the second training request also includes at least one of the following: a model identifier of the bilateral model, a model identifier of the first part, a model identifier of the second part, a first identifier of the network device, a second identifier of the terminal device, indication information for indicating the type of inference task for which the bilateral model will be used, or performance requirements of the bilateral model.

[0026] In some implementations, the method may further include: sending a request message to a second network management device, the request message including a first identifier of the network device; receiving a response message from the second network management device, the response message including at least one of the following: a computing power status of the network device, a storage status of the network device, energy consumption information of the network device, or a network energy consumption target value of the network device; and determining, based on the response message, at least one of the following: an energy consumption requirement for training the bilateral model, a first energy consumption requirement for an inference task to be used by the first part, a second energy consumption requirement for an inference task to be used by the second part, a first model size requirement for the first part, or a second model size requirement for the second part. In this way, the first network management device can determine the requirements for training based on the status information, thereby ensuring the availability of the trained model.

[0027] In some implementations, the method may further include: sending a second model training report to the server, the second model training report including training data used for training the second portion and structural information of the second portion. In this manner, the training data used for training the second portion may be further provided to the server, so that the server can train the second portion to obtain a matching portion.

[0028] In some implementations, the second model training report also includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, a first identification of the network device, an actual performance of the first part, an actual energy consumption of training the first part, an estimated energy consumption of the inference task for which the first part will be used, an actual model size of the first part, a performance requirement of the second part, a second energy consumption requirement of the inference task for which the second part will be used, or a second model size requirement of the second part.

[0029] In a second aspect of the present disclosure, a method for model training is provided. The method includes: sending a first training request to a second network management device, the first training request instructing to train a first part of a bilateral model, the bilateral model including the first part to be deployed at the network device and the second part to be deployed at the terminal device; and receiving a model training report from the second network management device, the model training report including at least a model identifier of the first part.

[0030] The method may be performed by a first network management device. In this manner, the first network management device may instruct a second network management device to train at least a first portion of a bilateral model, thereby supporting use of the bilateral model in an air interface.

[0031] In some implementations, the first training request includes training data for training the first portion and structural information of the first portion. In this manner, the second network management device can perform training based on the training data and structural information, thereby ensuring that the resulting first portion matches and correlates with the trained second portion.

[0032] In some implementations, the model training report includes a first model training report, which includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, an actual performance of the first part, an actual energy consumption of training the first part, an estimated energy consumption of an inference task for which the first part will be used, or an actual model size of the first part.

[0033] In some implementations, the method may further include receiving a second training request from the server, the second training request instructing to train a first part of the bilateral model, the second training request including training data for training the first part and structural information of the first part. In this manner, it can be determined that the server has completed training of the second part, and the second network management device can be instructed to train the second part.

[0034] In some implementations, the model training report includes a second model training report, which includes training data used to train the second part and structural information of the second part. In some implementations, the method may further include sending a training request to the server, where the training request includes training data used to train the second part and structural information of the second part. In this manner, the server can perform training to obtain the second part. Furthermore, because the training data and structural information used to train the second part come from the second network management device, the second part trained by the server will match the first part trained by the second network management device.

[0035] In some implementations, the second model training report further includes at least one of the following: a model identifier of the bilateral model, a model identifier of the first part, a first identifier of the network device, actual performance of the first part, actual energy consumption during training of the first part, estimated energy consumption of the inference task for which the first part will be used, and an actual model size of the first part. In this way, the first network management device can obtain information about the training process of the first part.

[0036] In some implementations, the method may further include: forwarding the second model training report to a server; receiving a notification message from the server indicating that the second part has been loaded onto the terminal device; and sending a model loading request to the second network management device, the model loading request instructing the second network management device to load the first part onto the network device. In this manner, after the server deploys the second part onto the terminal device, the second network management device may be instructed to deploy the first part onto the network device based on the model loading request from the server. This ensures the association between the first part and the second part, thereby ensuring the use of the bilateral model.

[0037] In some implementations, the method further includes: receiving a second training request from a server, the second training request requesting training of the first part of the bilateral model and also requesting training data for training the second part; and sending a first training request to a second network management device, the first training request requesting training of the first part of the bilateral model and also requesting training data for training the second part. Exemplarily, the method further includes: receiving a first model training report from the second network management device, the first model training report including training data for training the second part and structural information of the second part, the first model training report indicating that training of the first part has been completed. Exemplarily, the method further includes: sending a second model training report to the server, the second model training report including training data for training the second part and structural information of the second part, the second model training report indicating that training of the first part has been completed. In this way, the server can obtain the second part through training based on the second model training report, thereby supporting the use of the bilateral model over the air interface.

[0038] In a third aspect of the present disclosure, a method for model training is provided. The method includes: obtaining training data for training a second part of a bilateral model, where the bilateral model includes a first part to be deployed at a network device and a second part to be deployed at a terminal device; performing model training based on the training data to obtain the second part; and sending the second part to the terminal device.

[0039] The method can be executed by a server. In this way, the server can assist in the training process of the bilateral model of the air interface, specifically training the second part and deploying it on the terminal device, thereby supporting the use of the bilateral model in the air interface.

[0040] In some implementations, the method may further include sending a second training request to the first network management device, the second training request instructing to train a first portion of the bilateral model, the second training request including training data for training the first portion of the bilateral model and structural information of the first portion. In this manner, the training data and structural information for the first portion are provided, enabling the management domain to train the first portion to match the second portion.

[0041] In some implementations, the method may further include: receiving a first model training report from the first network management device, the first model training report indicating that the first portion has been trained. In this way, the server may determine that the first portion has been trained by the management domain based on the first model training report.

[0042] In some implementations, obtaining training data for training the second part of the bilateral model includes receiving a second model training report from the first network management device, the second model training report including training data for training the second part and structural information of the second part. In this manner, the training data and structural information for the second part provided by the management domain can be obtained, thereby enabling the trained second part to match the first part trained by the management domain.

[0043] In some implementations, the second model training report also includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, a first identification of the network device, the actual performance of the first part, the performance requirement of the second part, a second energy consumption requirement of the inference task for which the second part will be used, or a second model size requirement for the second part.

[0044] In some implementations, the method may further include: sending a notification message to the first network management device, the notification message indicating that the second part has been loaded onto the terminal device. In this way, the management domain can further deploy the first part to the network device based on the notification message, thereby ensuring the use of the bilateral model.

[0045] In a fourth aspect of the present disclosure, a system is provided. The system includes: a first network management device and a second network management device, wherein the first network management device can be configured to perform the operations of the method described in the second aspect or any implementation thereof, and the second network management device can be configured to perform the operations of the method performed by the second network management device described in the first aspect or any implementation thereof.

[0046] In some implementations, the system may further include a server, wherein the server may be configured to perform the operations of the method in the third aspect or any implementation as described above. For example, the server may obtain the second part of the bilateral model through training.

[0047] In some implementations, the system may further include a network device configured to receive a first portion of the bilateral model trained by the second network management device. In some implementations, the network device may further be configured to receive a second portion from the server and send the second portion to the terminal device. In some implementations, the system may further include a terminal device configured to receive the second portion from the network device.

[0048] In some other implementations, the second network management device may be further configured to obtain the second part of the bilateral model through training.

[0049] It should be noted that some implementation methods and beneficial effects of the aforementioned first to third aspects are also applicable to the fourth aspect. For the sake of brevity, they will not be repeated here.

[0050] In a fifth aspect of the present disclosure, a system is provided. The system includes: a first network management device and a server, wherein the first network management device can be configured to perform the operations of the method executable by the first network management device in the first aspect or any implementation described above, and the server can be configured to perform the operations of the method in the third aspect or any implementation described above. For example, the server can obtain the second part of the bilateral model through training.

[0051] In some implementations, the system may further include a network device configured to receive a first portion of the bilateral model trained by the first network management device. In some implementations, the network device may further be configured to receive a second portion from a server and send the second portion to a terminal device. In some implementations, the system may further include a terminal device configured to receive the second portion from the network device.

[0052] In some other implementations, the first network management device may be further configured to obtain the second part of the bilateral model through training.

[0053] It should be noted that some implementation methods and beneficial effects of the aforementioned first to third aspects are also applicable to the fifth aspect. For the sake of brevity, they will not be repeated here.

[0054] In a sixth aspect of the present disclosure, a method is provided. The method may include: a first network management device performing the operations of the method described in the second aspect or any implementation thereof, and a second network management device performing the operations of the method described in the first aspect or any implementation thereof that can be performed by the second network management device.

[0055] In some implementations, the method further includes: the server performing the operations of the method in the third aspect or any implementation as described above. For example, the server may obtain the second part of the bilateral model through training.

[0056] In some implementations, the method further includes: receiving, by the network device, the second portion from the server; and sending the second portion to the terminal device. In some implementations, the method further includes: receiving, by the terminal device, the second portion from the network device.

[0057] In some other implementations, the method further includes: obtaining the second part of the bilateral model through training by the second network management device.

[0058] In a seventh aspect of the present disclosure, a method is provided, which includes: a first network management device performing the operations of the method executable by the first network management device in the first aspect or any implementation manner described above.

[0059] In some implementations, the method further includes: executing, by the server, the operations of the method in the third aspect or any of the implementations described above. For example, the server may obtain the second part of the bilateral model through training.

[0060] In some implementations, the method further includes: receiving, by the network device, the second portion from the server; and sending the second portion to the terminal device. In some implementations, the method further includes: receiving, by the terminal device, the second portion from the network device.

[0061] In some other implementations, the method further includes: obtaining the second part of the bilateral model through training by the first network management device.

[0062] In an eighth aspect of the present disclosure, a network management device is provided, comprising a component for performing the operations according to the method of the first aspect or any implementation thereof, or comprising a component for performing the operations according to the method of the second aspect or any implementation thereof. Optionally, the component may be implemented as a unit, module, etc.

[0063] In a ninth aspect of the present disclosure, a server is provided, comprising a component for performing the operations according to the method in the third aspect or any implementation thereof. Optionally, the component may be implemented as a unit, a module, etc.

[0064] In the tenth aspect of the present disclosure, a network management device is provided, comprising a transceiver, a processor, and a memory, wherein the memory stores instructions executed by the processor, and when the instructions are executed by the processor, the network management device performs the operations of the method in the aforementioned first aspect, second aspect, or any implementation thereof.

[0065] In the eleventh aspect of the present disclosure, a server is provided, comprising a transceiver, a processor, and a memory, wherein the memory stores instructions executed by the processor, and when the instructions are executed by the processor, the server performs the operations of the method in the aforementioned third aspect or any of its implementations.

[0066] In a twelfth aspect of the present disclosure, a computer-readable storage medium, such as a non-transitory computer-readable storage medium, is provided. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the methods of the first to third aspects or any implementation thereof.

[0067] In a thirteenth aspect of the present disclosure, a chip or a chip system is provided, which includes a processing circuit configured to perform the operations according to the method of the first to third aspects or any implementation thereof.

[0068] In a fourteenth aspect of the present disclosure, a computer program or computer program product is provided. The computer program or computer program product is tangibly stored on a computer-readable medium and includes computer-executable instructions that, when executed, cause a device to perform the operations according to the method of the first to third aspects or any implementation thereof.

[0069] It should be noted that some embodiments and beneficial effects of the aforementioned method aspects are also applicable to device aspects, user equipment aspects, network equipment aspects, computer-readable storage media aspects, chips or chip systems aspects, computer programs or computer program products aspects, and for the sake of brevity, they will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The above and other features, advantages and aspects of various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings.

[0071] In the accompanying drawings, the same or similar reference numerals indicate the same or similar elements, wherein:

[0072] FIG1 shows a schematic diagram of a bilateral AI model for channel state information (CSI) feedback;

[0073] FIG2 shows a schematic structural diagram of a network system in which some embodiments of the present disclosure can be applied;

[0074] FIG3 shows a schematic structural diagram of another network system to which some embodiments of the present disclosure can be applied;

[0075] FIG4 shows a signaling diagram according to some embodiments of the present disclosure;

[0076] FIG5 shows another signaling diagram according to some embodiments of the present disclosure;

[0077] FIG6 shows another signaling diagram according to some embodiments of the present disclosure;

[0078] FIG7 shows yet another signaling diagram according to some embodiments of the present disclosure;

[0079] FIG8 shows another signaling diagram according to some embodiments of the present disclosure;

[0080] FIG9 shows another signaling diagram according to some embodiments of the present disclosure;

[0081] 10A-10C are schematic diagrams showing training and inference scenarios according to some embodiments of the present disclosure;

[0082] FIG11 shows a schematic block diagram of a network management device according to some embodiments of the present disclosure;

[0083] FIG12 shows a schematic block diagram of a network management device according to some embodiments of the present disclosure;

[0084] FIG13 shows a schematic block diagram of a server according to some embodiments of the present disclosure; and

[0085] FIG14 shows a schematic block diagram of an example device that may be used to implement embodiments of the present disclosure. DETAILED DESCRIPTION

[0086] The following describes embodiments of the present disclosure in more detail with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0087] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to." The term "based on" should be understood as "based at least in part on." The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment." The terms "first," "second," etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0088] The embodiments of the present disclosure may be implemented according to any suitable communication protocol, including but not limited to third generation (3G), fourth generation (4G), fifth generation (5G), sixth generation (6G) and other cellular communication protocols, wireless local area network communication protocols such as Institute of Electrical and Electronics Engineers (IEEE) 802.11, and / or any other protocol currently known or developed in the future.

[0089] The technical solutions of the embodiments of the present disclosure are applied to communication systems that follow any appropriate communication protocols, such as: Universal Mobile Telecommunications Service (UMTS), Long Term Evolution (LTE) system, Wideband Code Division Multiple Access (WCDMA), Code Division Multiple Access 2000 (CDMA2000), Time Division-Synchronization Code Division Multiple Access (TD-SCDMA), Frequency Division Duplex (FDD) system, Time Division Duplex (TDD), fifth generation (5G) system or new radio (NR), future evolved 6G and 7G communication systems, and the like.

[0090] It should be understood that the embodiments of the present disclosure can be applied to any communication system having similar problems, such as a wireless local area network (WLAN), a wired communication system, or other communication systems developed in the future.

[0091] The term "terminal device" used in this disclosure refers to any terminal device that can perform wired or wireless communication with network devices or with each other. A terminal device may sometimes be referred to as a user equipment (UE). A terminal device may be any type of mobile terminal, fixed terminal, or portable terminal. As an example, a terminal device may include a mobile phone, a site, a unit, a device, a mobile terminal (MT), a subscription station, a portable subscription station, an Internet node, a communicator, a desktop computer, a laptop computer, a notebook computer, a tablet computer, a personal communication system device, a personal navigation device, a personal digital assistant (PDA), a positioning device, a radio broadcast receiver, an e-book device, a gaming device, an Internet of Things (IoT) device, a vehicle-mounted device, an aircraft, a virtual reality (VR) device, an augmented reality (AR) device, a wearable device, a terminal device in a 5G network or any terminal device in an evolved public land mobile network (PLMN), other devices that can be used for communication, or any combination thereof. The embodiments of the present disclosure are not limited to this.

[0092] The term "network device" as used in this disclosure refers to an entity or node that can be used to communicate with a terminal device, for example, an access network network device (i.e., an access network device). An access network device can be a device deployed in a RAN to provide wireless communication functions for a mobile terminal, for example, a RAN network device. Access network devices may include various types of base stations (BS). As an example, access network devices may include various forms of macro base stations, micro base stations, pico base stations, femto base stations, relay stations, access points, satellites, remote radio units (RRUs), radio heads (RHs), remote radio heads (RRHs), etc., and may also be devices that perform base station functions in device-to-device (D2D) communication and machine-to-machine (M2M) communication. In systems using different wireless access technologies, the names of access network equipment may vary. For example, in Long Term Evolution (LTE) networks, it is called an evolved NodeB (eNB or eNodeB); in 3G networks, it is called a NodeB (NB); in 5G networks, it may be called a gNodeB (gNB) or a NR NodeB (NR NB), and so on. In some scenarios, access network equipment may include a Central Unit (CU) and / or a Distributed Unit (DU). The CU and DU can be located in different locations, for example: the DU can be remotely located in a high-traffic area, while the CU can be located in a central computer room. Alternatively, the CU and DU can be located in the same computer room. The CU and DU can also be different components within the same rack. Access network equipment can also be devices deployed in an open RAN (O-RAN) to provide wireless communication functions for mobile terminals. For example, the access network equipment may include an open CU (O-CU) and / or an open DU (O-DU). For the convenience of description, in the subsequent embodiments of the present disclosure, the above-mentioned devices providing wireless communication functions for mobile terminals are collectively referred to as network devices, and the embodiments of the present disclosure are no longer specifically limited.

[0093] As mentioned above, the AI ​​model used over the air interface (i.e., the RAN domain AI model) can include unilateral AI models (referred to as unilateral models) and bilateral AI models (referred to as bilateral models). For example, unilateral AI models can be used for beam management and positioning, while bilateral AI models can be used for CSI feedback. Characteristics of bilateral AI models include: the model consists of two parts, with the output of the first part serving as the input of the second part; model training requires collaborative work to ensure that the original input and final output are as close as possible; and model inference requires collaborative work, with the models used by both parties performing the inference matching each other.

[0094] Taking CSI feedback as an example, Figure 1 shows a schematic diagram of a bilateral AI model 100 for CSI feedback. In Figure 1, a decoder model deployed at an access network device (e.g., BS) 110 and an encoder model deployed at a terminal device (e.g., UE) 120 are shown. In a frequency division multiplexing network, the BS needs to obtain downlink CSI feedback to perform precoding to achieve performance gains. However, the UE consumes more air interface resources to transmit the original CSI. Therefore, in order to reduce the overhead of CSI feedback, a bilateral AI model can be used to implement the CSI compression and reconstruction mechanism. Specifically, a bilateral AI model can be constructed by using a neural network structure of an autoencoder (AutoEncoder), which consists of an encoder model and a decoder model. The encoder model is deployed on the UE side to convert the original CSI matrix (expressed as X t ) Use neural network to perform dimensionality reduction processing and obtain compressed CSI (expressed as Y t ); The decoder model is deployed on the BS side to recover the received compressed CSI using a neural network to obtain the estimated original CSI (expressed as ).

[0095] Model lifecycle management involves model training, model testing, model simulation, model loading, and model inference. Model lifecycle management can be performed by the management domain, with specific model training and model inference performed by the management domain, the RAN domain, and / or the core network domain. For bilateral AI models inferred in the RAN domain, the management domain can assist in model training and then deploy the models to access network devices and terminal devices for model inference. However, the details of how the management domain specifically assists in bilateral AI model training in the RAN domain require further study.

[0096] FIG2 shows a schematic structural diagram of a network system 200 in which some embodiments of the present disclosure can be applied. FIG2 includes a 3GPP network domain 250 and a server 290 outside of the 3GPP network domain 250. The 3GPP network domain 250 includes an NMS 210, an EMS 220, an access network device 230, and a terminal device 240, wherein the NMS 210 and / or the EMS 220 perform model training functions, while the access network device 230 and the terminal device 240 perform model inference functions. Server 290 can be a UE server, which can be a server provided by the manufacturer of the terminal device 240 or a server provided by the chip manufacturer within the terminal device 240. For example, server 290 can perform model training functions, provide software download or upgrade services for the terminal device 240, and so on.

[0097] FIG3 shows a schematic diagram of a network system 300 in which some embodiments of the present disclosure can be applied. System 300 is based on the O-RAN architecture, where O-RAN is designed to implement an intelligent, open access network. O-RAN is a concept proposed based on RAN interoperability and standardization, which provides the same interconnection standards for white box hardware and software from different vendors. As shown in FIG3 , system 300 includes a service management and orchestration framework (SMO) 305, which includes a non-real-time RAN intelligent controller (non-RT RIC or NRT RIC) 310. System 300 also includes a near-real-time RAN intelligent controller (near-RT RIC or nRT RIC) 320, access network equipment 330, terminal equipment 340, and server 390, where access network equipment 330 includes an O-CU 332, an O-DU 334, and an O-RU 336. The non-RT RIC 310 and / or the near-RT RIC 320 perform model training functions, while the O-CU 332, the O-DU 334, and the terminal device 340 perform model inference functions. The server 390 may be a UE server, which may be provided by the manufacturer of the terminal device 340 or provided by the chip manufacturer within the terminal device 340. For example, the server 390 may perform model training functions and provide software download or upgrade services for the terminal device 340.

[0098] Embodiments of the present disclosure involve a management domain performing model training for a bilateral AI model, where the management domain may include the NMS 210 and EMS 220 as shown in FIG2 , or the management domain may include the non-RT RIC 310 and near-RT RIC 320 as shown in FIG3 . It is understood that the management domain that performs model training may also be referred to as a management plane, and may include other devices, network elements, or systems in other scenarios or architectures, which are not listed one by one in this disclosure.

[0099] In the following detailed embodiments, for ease of description, the system 200 shown in FIG2 will be used for illustration. However, it should be understood that the following embodiments can also be applied to the system 300 shown in FIG3 . For example, the NMS 210 can be replaced with a non-RT RIC 310, the EMS 220 can be replaced with a near-RT RIC 320, the access network device 230 can be replaced with an access network device 330 (or replaced with an O-CU 332 and an O-DU 334), the terminal device 240 can be replaced with a terminal device 250, and the server 290 can be replaced with a server 390, so that the following embodiments can be applied to the system 300. In addition, the following example embodiments can also be applied to other systems or architectures not shown herein, which are not listed here one by one.

[0100] In an embodiment of the present disclosure, the bilateral AI model may include a first part deployed on an access network device and a second part deployed on a terminal device, and the first part and the second part are matched (or associated). Exemplarily, the matching first and second parts can be understood as follows: when the output of the second part is used as the input of the first part, the output of the first part can be approximated to the input of the second part. For the convenience of description, in the following embodiments, it is assumed that the first part is a decoder model and the second part is an encoder model. This will be explained in more detail below in conjunction with Figures 4 to 10C.

[0101] Figure 4 shows a schematic diagram of a signaling interaction 400 according to some embodiments of the present disclosure. Figure 4 involves a server 290, an NMS 210, an EMS 220, an access network device 230, and a terminal device 240. It is understandable that although the signaling interaction 400 is described in conjunction with Figure 2, the process can also be applied to other systems, and the present disclosure is not limited to this. In Figure 4, the server 290 and the EMS 220 perform model training. Specifically, the server 290 obtains the part of the bilateral AI model to be deployed to the terminal device 240 (referred to as the encoder model or UE-side model), and the EMS 220 obtains the part of the bilateral AI model to be deployed to the access network device 230 (referred to as the decoder model or BS-side model).

[0102] At 401, the NMS 210 may, for example, open AI / ML management services to the server 290 through a management service open interface. For example, the AI / ML management services may include model training, model deployment notifications, etc. It is understandable that since the server 290 does not belong to the 3GPP network domain and cannot directly call services in the 3GPP network, the NMS 210 may provide services to the server 290 by opening the network capability. It should be noted that step 401 is a prerequisite step of this embodiment, and the present disclosure does not limit the specific implementation of the network capability opening method. In this way, the interface of the model training service and the model loading service for the bilateral AI model can be opened to the server 290 outside the 3GPP network domain to ensure the normal execution of the model training.

[0103] At 402, server 290 obtains training data. Specifically, terminal device 240 may perform operations such as data collection, measurement, and calculation to obtain training data, and terminal device 240 may report the training data to server 290. For example, terminal device 240 may provide the training data to server 290 via a proprietary interface, such as the Hypertext Transport Protocol (HTTP) communication protocol. Assuming that this embodiment is intended to train a bilateral AI model for CSI feedback, the training data may include raw CSI measured by terminal device 240.

[0104] At 403, server 290 performs model training based on the training data to obtain an encoder model (i.e., a UE-side model) and an initial decoder model. The model training process may involve multiple rounds of iterations. Details of the model training may be referenced to existing technologies and are not limited in this disclosure.

[0105] At 404, the server 290 sends a first model training request to the NMS 210. The first model training request includes: training data for training the decoder model (i.e., the BS-side model) and structural information of the decoder model. The first model training request may also include one or more of the following: a model identifier of the bilateral AI model, an identifier of the access network device 230, an identifier of the terminal device group, a model identifier of the encoder model (i.e., the UE-side model), an indication of the type of inference task for which the bilateral AI model will be used, model performance requirements (such as performance requirements of the decoder model), bilateral model training instructions, etc.

[0106] Exemplarily, after completing model training at 403, the server 290 may construct training data for training a decoder model (i.e., a BS-side model) by the management domain based on the training data obtained at 402, and optionally based on the initial decoder model obtained by the model training. Exemplarily, the first model training request may directly include training data for training the decoder model (i.e., a BS-side model). Exemplarily, the server 290 may store training data for training the decoder model (i.e., a BS-side model), and the first model training request may include the address of the storage to indicate the training data for training the decoder model (i.e., a BS-side model). Thereafter, the management domain (e.g., NMS 210 or EMS 220) may obtain training data for training the decoder model (i.e., a BS-side model) based on the storage address. For example, the management domain (e.g., NMS 210 or EMS 220) may download training data for training the decoder model (i.e., a BS-side model) from the storage address via a File Transfer Protocol (FTP).

[0107] The first model training request includes the structural information of the decoder model. By indicating the structural information, it is possible to ensure that the decoder model trained by the management domain (such as NMS 210 or EMS 220) matches (or connects) the encoder model trained by the server 290, thereby ensuring the availability of the bilateral AI model. Exemplarily, the structural information of the decoder model may include one or more of the following: the network type of the decoder model, the number of network layers, the number of nodes in the input layer, the number of nodes in the output layer, etc., where the network type may be a convolutional neural network, a fully connected neural network, a recurrent neural network, etc. Optionally, the first model training request may include an initial decoder model to implicitly indicate the structural information of the decoder model. In this way, the management domain (such as EMS 220) can perform training (such as in 408) based on the initial decoder model, which can reduce the computational complexity of training and improve the efficiency of training.

[0108] The first model training request may also include a model identifier of the bilateral AI model and / or a model identifier of the encoder model (i.e., the UE-side model). In some examples, the model identifier of the encoder model (i.e., the UE-side model) may uniquely represent or indicate the encoder model trained by the server 290. In some examples, the model identifier of the bilateral AI model may be used to uniquely identify the bilateral AI model. Optionally, the model identifier of the bilateral AI model may be the same as the model identifier of the decoder model to be deployed at the access network device 230. For example, if a decoder model has been previously trained and a model identifier of the decoder model has been assigned, the bilateral AI model may be indicated by the model identifier of the decoder model, and the previously trained decoder model may be instructed to be retrained. Optionally, the model identifier of the bilateral AI model may be a unique identifier generated based on a combination of the model identifier of the encoder model and the model identifier of the decoder model.

[0109] The first model training request may also indicate the type of reasoning task for which the bilateral AI model will be used. For example, the first model training request includes indicative information about the type of reasoning task for which the bilateral AI model will be used. Exemplarily, the type of reasoning task, i.e., the reasoning type, may indicate the actual use case for which the bilateral AI model will be used. For example, the reasoning type may be CSI feedback, beam management, etc. Optionally, if the bilateral AI model is being trained for the first time or the model identifier of the bilateral AI model has not yet been assigned (unknown), the model to be trained may be indicated by the reasoning type.

[0110] The first model training request may also include an identifier of the access network device 230 (i.e., a base station identifier) ​​and / or an identifier of the terminal device group (i.e., a UE group identifier). The identifier of the access network device 230 (i.e., a base station identifier) ​​can be used to indicate the access network device to which the decoder model can be applied. For example, the base station identifier can be implemented as a network unique identifier name (Distinguished Name, DN) or a geographical area indication. The identifier of the terminal device group (i.e., a UE group identifier) ​​can be used to indicate the terminal devices to which the encoder model trained by the server 290 can be applied. For example, the terminal device group includes at least the terminal device 240 shown in the figure. For example, the UE group identifier can be implemented as the manufacturer name, software version, chip manufacturer name, or chip model of the UE.

[0111] The first model training request may also include a bilateral model training indication, which may indicate that a bilateral AI model needs to be trained. In some examples, the bilateral model training indication may be represented as a Boolean value, for example, "true" indicates that a bilateral AI model needs to be trained, while "false" indicates that a bilateral AI model does not need to be trained (i.e., a unilateral AI model is trained). In other examples, the bilateral model training indication may be represented as a coded value, for example, "1" indicates that a bilateral AI model needs to be trained, while "0" indicates that a bilateral AI model does not need to be trained (i.e., a unilateral AI model is trained). It is understandable that the bilateral model training indication may be represented in other forms, and the present disclosure is not limited to this. In some embodiments, the bilateral model training indication may be implicitly indicated. For example, the first model training request includes training data for training the decoder model (i.e., the BS-side model) and structural information of the decoder model. Then, NMS 210 may know that the decoder model needs to be trained, but NMS 210 may not know whether the decoder model to be trained is a unilateral AI model or a BS-side model in the bilateral AI model. Furthermore, the first model training request may also include the model identifier of the encoder model that has been trained by the server 290. Then, NMS 210 may determine that the decoder model to be trained is to be matched with the encoder model that has been trained by the server 290, so that NMS 210 can determine that the current training is for the bilateral AI model. In other words, it can be considered that the first model training request implicitly indicates the bilateral model training indication.

[0112] The first model training request may optionally further indicate a performance requirement for the decoder model. For example, the performance requirement may indicate an accuracy requirement for the decoder model when it is subsequently used for an inference task, such as 90%, 85%, or other values.

[0113] Optionally, at 405, NMS 210 sends a device query request message to EMS 220; and at 406, EMS 220 sends a device query response message to NMS 210. For example, the device query request message may be used to request status information of an access network device, e.g., the device query request message may be referred to as a device status query request message. Accordingly, the device query response message may be used to provide status information of the access network device being queried, e.g., the device query response message may be referred to as a device status query response message.

[0114] In some embodiments, the status information of the access network device may include one or more of the following: computing power status, storage status, energy consumption status, and network energy consumption target value of the access network device. Exemplarily, the device query response message may include computing resource utilization to indicate the computing power status. Exemplarily, the device query response message may include storage resource utilization to indicate the storage status. Exemplarily, the device query response message may include the current energy consumption or energy efficiency value to indicate the energy consumption status. Exemplarily, the device query response message may include the network's target value for energy consumption to indicate the network's target value for energy consumption, for example, the network's target value for energy consumption is: the overall energy consumption within a certain geographical range is less than 500 kilojoules (kJ).

[0115] In this way, through 405 and 406, NMS 210 can obtain the status information of the access network device, and then NMS 210 can determine the requirements for training based on the status information, for example, the requirements can be indicated to EMS 220 in the following second model training request, thereby ensuring the availability of the first part of the trained bilateral AI model (i.e., the decoder model) at the access network device, thereby supporting the use of the bilateral AI model.

[0116] At 407, NMS 210 sends a second model training request to EMS 220. The second model training request may include part or all of the first model training request described above. The second model training request includes training data for training the decoder model (i.e., the BS-side model) and structural information of the decoder model. In this way, the training data and structural information for training the decoder model can be provided to EMS 220, allowing EMS 220 to obtain a decoder model based on this training, thereby ensuring that it matches the trained encoder model.

[0117] The first model training request may also include one or more of the following: the model identifier of the bilateral AI model, the identifier of the access network device 230, the identifier of the terminal device group (including the terminal device 240), the model identifier of the encoder model (i.e., the UE-side model) that has been trained by the server 290, indication information of the type of inference task for which the bilateral AI model will be used, model performance requirements, bilateral model training instructions, etc.

[0118] In some embodiments, the second model training request may also include one or more of the following regarding the decoder model: training energy consumption requirement, inference energy consumption requirement, model size requirement, or lightweight indication. Specifically, NMS 210 may determine one or more of the following based on the device query response message received at 406: training energy consumption requirement, inference energy consumption requirement, model size requirement, or lightweight indication. For example, the computing power status and storage status are used as a reference to determine the computing and storage requirements for bilateral AI model training and inference, while the energy consumption status and network energy consumption target value are used as a reference to determine the energy requirements for bilateral AI model training and inference. For example, the training (or inference) energy consumption requirement may include the amount of energy consumed during the training (inference) process (e.g., duration, power consumption). For example, the model size requirement may indicate the maximum required (or upper limit) storage space occupied by the trained model, e.g., 100MB. The lightweight indication indicates whether the model needs to be lightweighted, and the lightweighting method used, such as pruning, quantization, or distillation. Model pruning can also be called model sparsification. For example, some model weight parameters can be directly deleted. Model quantization can compress the model weight parameters. Model distillation compresses the model based on knowledge distillation.

[0119] At 408, EMS 220 performs model training based on the second model training request. Specifically, by using the training data for training the decoder model (i.e., the BS-side model) indicated in the second model training request and the structural information of the decoder model, EMS 220 obtains the decoder model (i.e., the BS-side model) through training.

[0120] Optionally, at 409, EMS 220 performs lightweight processing on the decoder model. In some examples, if the second model training request includes a lightweight indication, indicating that lightweight processing is required, EMS 220 performs lightweight processing. In other examples, even if the second model training request does not include a lightweight indication, or the second model training request indicates that lightweight processing is not required, EMS 220 may decide to perform lightweight processing on its own. For example, EMS 220 may decide to perform lightweight processing based on model performance requirements, training energy consumption requirements, inference energy consumption requirements, model size requirements, etc. of the decoder model. Exemplarily, EMS 220 may perform quantization, pruning, and other processing on the decoder model to obtain a decoder model after lightweight processing. It can be understood that lightweight processing is to compress the model under the premise of meeting the model performance requirements (such as accuracy) as much as possible to reduce the computational complexity and size (such as the number of parameters) of the model.

[0121] In this way, EMS 220 can obtain a trained decoder model that matches the encoder model trained by server 290 through 408, or through 408 and 409. It should be noted that although 409 is shown as a step independent of 408 in Figure 4, in other embodiments, 408 may include 409, that is, lightweight processing can be performed during the model training process. It can be understood that since the training at 408 and 409 is based on the training data and structural information provided by the server for training the decoder model, the trained decoder model can match the encoder model trained by the server. Since the various requirements in the second model training request are taken into account at 408 and 409, it is ensured that the trained decoder model can adapt to the status information at the access network device, ensuring the availability of the decoder model.

[0122] At 410, EMS 220 sends a second model training report to NMS 210, indicating that the decoder model has been trained. For example, the second model training report may include the model identifier of the trained decoder model and the identifier of access network device 230. For example, the second model training report may also include one or more of the following: actual performance of the decoder model, actual training energy consumption of the decoder model, actual inference energy consumption of the decoder model, and actual model size of the decoder model.

[0123] At 411, NMS 210 sends a first model training report to server 290 to indicate that the decoder model has been trained. For example, the first model training report may include some or all of the information in the second model training report. For example, the first model training report may include: a model identifier of the trained decoder model and the actual performance of the decoder model.

[0124] At 412, the server 290 sends the encoder model to the terminal device 240. In some examples, the encoder model can be expressed in the form of model parameters or carried in the form of a model file. Specifically, the server 290 can send a model sending request to the terminal device 240, and the model sending request includes the encoder model itself, or includes model parameters or the file address of the model file. Exemplarily, the model sending request may also include a bilateral model indication, a model identifier of the encoder model, and a model identifier of the decoder model. In this way, the terminal device 240 can directly obtain the encoder model, or determine the encoder model through the model parameters, or obtain the encoder model from the file address of the model file. The terminal device 240 can determine that the encoder model is the UE-side model in the bilateral AI model based on the bilateral model indication. In addition, the terminal device 240 can determine which decoder model matches the encoder model based on the model identifier of the encoder model and the model identifier of the decoder model. In this way, the terminal device 240 can use the encoder model from the server 290 as the UE-side model of the bilateral AI model instead of using it as a unilateral model, thus ensuring the association with the decoder model at the access network device 230.

[0125] At 413, server 290 sends a model delivery notification to NMS 210 to indicate that the encoder model has been deployed at terminal device 240. For example, the model delivery notification may include a bilateral model indication and a model identifier of the encoder model, indicating that the encoder model has been delivered.

[0126] At 414, NMS 210 sends a model loading request to EMS 220, and then at 415, EMS 220 loads the decoder model to access network device 230. Exemplarily, the model loading request may include a model identifier of the decoder model and an identifier of access network device 230, instructing EMS 220 to deploy the decoder model to access network device 230. Exemplarily, the model loading request may also include a bilateral model indication, a model identifier of the encoder model, and a terminal device group identifier. In this way, access network device 230 can use the decoder model from EMS 220 as a BS-side model of the bilateral AI model rather than as a unilateral model, thereby ensuring association with the encoder model at terminal device 240.

[0127] In some examples, the decoder model can be expressed in the form of model parameters or carried in the form of a model file. For example, EMS 220 can send another model loading request to access network device 230. This other model loading request can include the decoder model itself, or include model parameters of the decoder model or the file address of the model file of the decoder model. In this way, access network device 230 can directly obtain the decoder model, or determine the decoder model through model parameters, or obtain the decoder model from the file address of the model file. For example, the other model loading request can also include part or all of the model loading request described above, for example, the other model loading request includes a bilateral model indication, a model identifier of the encoder model, and a terminal device group identifier. In this way, access network device 230 can determine that the decoder model is the BS-side model in the bilateral AI model based on the bilateral model indication. Furthermore, access network device 230 can determine which encoder model matches the decoder model and which terminal device(s) the encoder model is deployed on based on the model identifier of the encoder model and the terminal device group identifier. For example, the bilateral model indication, the model identifier of the encoder model, and the terminal device group identifier can be understood as bilateral model association information to determine which encoder model matches the decoder model and which / those terminal devices the encoder model is deployed on.

[0128] In this way, through 401 to 415, the server 290 can be supported to call the model management service, and then the server 290 and EMS 220 can collaboratively complete the model training of the bilateral AI model, and are respectively deployed at the terminal device 240 and the access network device 230. In some embodiments, during the training process, the status information of the access network device 230 can be simultaneously considered, so that an appropriate model can be trained based on the computing power, storage, and energy consumption requirements of the access network device 230.

[0129] Furthermore, at 416, the bilateral AI model may be used in a subsequent model inference process. As an example, the model inference process for CSI feedback may refer to the description above in conjunction with FIG1.

[0130] By combining the signaling interaction 400 described in FIG4 , the service types (i.e., model training service, model loading service) opened by the management plane (such as NMS 210) can be increased, and the server 290 outside the 3GPP network domain 250 can be supported to call the model training service and model loading service of the management plane. The first / second model training request may include a bilateral model indication, a model identifier of the UE-side model, a model structure of the BS-side model, and training data of the BS-side model, thereby supporting the management plane (such as EMS 220) to train the BS-side model with the training data provided by the external server 290. The model loading request / another model loading request may include a bilateral model indication and / or bilateral model association information, thereby indicating that the UE-side model has been deployed, thereby supporting the deployment and reasoning of the bilateral AI model.

[0131] Figure 5 shows a schematic diagram of another signaling interaction 500 according to some embodiments of the present disclosure. Figure 5 involves a server 290, an NMS 210, an EMS 220, an access network device 230, and a terminal device 240. It is understandable that although the signaling interaction 500 is described in conjunction with Figure 2, the process can also be applied to other systems, and the present disclosure is not limited to this. In Figure 5, the server 290 and the EMS 220 perform model training. Specifically, the server 290 obtains the part of the bilateral AI model to be deployed to the terminal device 240 (referred to as the encoder model or UE-side model), and the EMS 220 obtains the part of the bilateral AI model to be deployed to the access network device 230 (referred to as the decoder model or BS-side model).

[0132] At 501, the NMS 210 may open the AI / ML management service to the server 290. This process is similar to 401 in FIG. 4 , and will not be described again to avoid repetition.

[0133] At 502, server 290 sends a third model training request to NMS 210. The third model training request may include an indication of training data for requesting UE-side model training. The third model training request may also include one or more of the following: a model identifier of the bilateral AI model, an identifier of access network device 230, an identifier of a terminal device group (including terminal device 240), an indication of the type of inference task for which the bilateral AI model will be used, model performance requirements (such as performance requirements of a decoder model), and a bilateral model training indication.

[0134] In some examples, the indication of training data for requesting UE-side model training can be represented as a Boolean value, for example, "true" indicates that training data for training the UE-side model is required, while "false" indicates that training data for training the UE-side model is not required. In other examples, the indication of training data for requesting UE-side model training can be represented as a coded value, for example, "1" indicates that training data for training the UE-side model is required, while "0" indicates that training data for training the UE-side model is not required.

[0135] In some embodiments, the bilateral model training indication and the indication of training data for requesting UE-side model training can be jointly indicated by a Boolean value or encoding. For example, "true" indicates that the bilateral AI model needs to be trained and training data for training the UE-side model is required. For example, "false" indicates that the bilateral AI model does not need to be trained. For example, the encoded value "00" indicates that the bilateral AI model does not need to be trained; "10" indicates that only the BS-side model needs to be trained; and "11" indicates that the bilateral AI model needs to be trained and training data for training the UE-side model is required.

[0136] For example, regarding the model identification of the bilateral AI model, the identification of the access network device 230, the identification of the terminal device group (including the terminal device 240), the indication information of the type of reasoning task for which the bilateral AI model will be used, the model performance requirements, the bilateral model training instructions, etc., refer to the terms described in the embodiment described above in combination with Figure 4, and will not be repeated here.

[0137] Optionally, at 503, NMS 210 sends a device query request message to EMS 220, and at 504, EMS 220 sends a device query response message to NMS 210. Regarding 503 and 504, reference may be made to 405 and 406 described above in conjunction with FIG. 4 and will not be repeated here.

[0138] At 505, NMS 210 sends a fourth model training request to EMS 220. The fourth model training request may include part or all of the third model training request described above. The fourth model training request includes an indication of requesting training data for UE-side model training. The fourth model training request may also include one or more of the following: a model identifier of the bilateral AI model, an identifier of the access network device 230, an identifier of the terminal device group (including terminal device 240), an indication of the type of inference task for which the bilateral AI model will be used, model performance requirements, and a bilateral model training instruction.

[0139] In some embodiments, the fourth model training request may further include one or more of the following regarding the decoder model: training energy consumption requirement, inference energy consumption requirement, model size requirement, or lightweight indication, etc. As described in detail above in conjunction with the embodiment of the second model training request in FIG4 .

[0140] At 506, EMS 220 obtains training data. Specifically, EMS 220 obtains the training data collected by terminal device 240 through access network device 230. Specifically, terminal device 240 may perform operations such as data collection, measurement, and calculation to obtain training data, and terminal device 240 may send the training data to access network device 230 via an uplink, which access network device 230 may then provide to EMS 220 (e.g., based on an instruction or request from EMS 220). Assuming that this embodiment is intended to train a bilateral AI model for CSI feedback, the training data may include raw CSI measured by terminal device 240.

[0141] At 507 , the EMS 220 performs model training using the training data based on the fourth model training request to obtain a decoder model (ie, a BS-side model) and an initial encoder model (ie, a UE-side model).

[0142] Optionally, at 508, EMS 220 may perform lightweight quantization on the decoder model. In some examples, if the fourth model training request includes a lightweight quantization indication indicating that lightweight quantization is required, EMS 220 performs lightweight quantization. In other examples, even if the fourth model training request does not include a lightweight quantization indication, or even if the fourth model training request indicates that lightweight quantization is not required, EMS 220 may decide to perform lightweight quantization at its discretion, as described in conjunction with 409 of FIG. 4 .

[0143] In this way, the EMS 220 can obtain a trained decoder model through 507, or through 507 and 508. It should be noted that although 508 is shown as a step independent of 507 in FIG5 , in other embodiments, 507 may include 508, that is, lightweight processing may be performed during the model training process.

[0144] At 509, EMS 220 sends a fourth model training report to NMS 210. Exemplarily, the fourth model training report may indicate that the decoder model has been trained, and the fourth model training report includes training data for training the encoder model (ie, UE-side model) and structural information of the encoder model.

[0145] Exemplarily, after the EMS 220 completes the model training at 507, it may construct training data for training the encoder model (i.e., the UE-side model) by the server 290 based on the training data obtained at 506 and, optionally, based on the initial encoder model obtained by the model training. Exemplarily, the fourth model training report may directly include the training data for training the encoder model (i.e., the UE-side model). Exemplarily, the fourth model training report may include a storage address of the training data for training the encoder model (i.e., the UE-side model), so that the server 290 may download the training data for training the encoder model (i.e., the UE-side model) from the storage address via FTP.

[0146] Exemplarily, the fourth model training report may include a model identifier of the trained decoder model and an identifier of the access network device 230. Exemplarily, the fourth model training report may also include one or more of the following: actual performance of the decoder model, actual training energy consumption of the decoder model, actual inference energy consumption of the decoder model, and actual model size of the decoder model.

[0147] At 510, NMS 210 sends a third model training report to server 290. Exemplarily, the third model training report may indicate that the decoder model has been trained, and the third model training report includes training data for training the encoder model (i.e., the UE-side model) and structural information of the encoder model. In this way, the training data and structural information for training the encoder model can be provided to server 290, so that server 290 obtains the encoder model based on this training, thereby ensuring matching with the trained decoder model. Exemplarily, the third model training report may include part or all of the fourth model training report. For example, the third model training report may include: a model identifier of the trained decoder model, an identifier of the access network device 230, and the actual performance of the decoder model.

[0148] Optionally, the third model training report may further include a model performance requirement of the encoder model (i.e., the UE-side model). For example, the model performance requirement of the encoder model indicates the accuracy requirement of the encoder model when it is subsequently used for inference tasks, such as 90%, 88%, or other values.

[0149] At 511, the server 290 performs model training based on the third model training report to obtain a trained encoder model. Specifically, the server 290 can construct a model based on the structural information of the encoder model, and perform model training based on the training data for training the encoder model (i.e., the UE-side model). It can be understood that since the training at 511 is based on the training data and structural information for training the encoder model provided by EMS 220, the trained encoder model can match the decoder model trained by EMS 220. In some examples, during the training process of 511, the model performance requirements for the encoder model in the third model training report can also be considered to ensure that the trained encoder model can adapt to the status information at the terminal device and ensure the availability of the encoder model.

[0150] At 512, server 290 sends the encoder model to terminal device 240. At 513, server 290 sends a model delivery notification to NMS 210, indicating that the encoder model has been deployed on terminal device 240. At 514, NMS 210 sends a model loading request to EMS 220, and at 515, EMS 220 loads the decoder model to access network device 230. At 516, the bilateral AI model can be used for subsequent model inference processes. Specifically, steps 512 to 516 can be referred to as steps 412 to 416 described in conjunction with FIG. For the sake of brevity, they will not be repeated here.

[0151] By combining the signaling interaction 500 described in FIG5 , the server 290 can be supported to call the model management service, and then the server 290 and the EMS 220 can collaboratively complete the model training of the bilateral AI model, and are deployed at the terminal device 240 and the access network device 230, respectively. In some embodiments, during the training process, the status information of the access network device 230 can be considered at the same time, so that the appropriate model can be trained based on the computing power, storage, and energy consumption requirements of the access network device 230. The service types (i.e., model training service, model loading service) opened by the management plane (such as NMS 210) can be increased to support the server 290 outside the 3GPP network domain 250 to call the model training service and model loading service of the management plane. The third / fourth model training request can include a bilateral model indication and an indication of training data for requesting UE-side model training, and the third / fourth model training report includes training data for training the encoder model and structural information of the encoder model, thereby supporting the server 290 outside the 3GPP network domain 250 to use the training data provided by the management plane to train the UE-side model. The model loading request / another model loading request may include a bilateral model indication and / or bilateral model association information, thereby indicating that the UE-side model has been deployed, thereby supporting the deployment and reasoning of the bilateral AI model.

[0152] Figure 6 shows a schematic diagram of another signaling interaction 600 according to some embodiments of the present disclosure. Figure 6 involves NMS 210, EMS 220, access network device 230 and terminal device 240. It can be understood that although the signaling interaction 600 is described in conjunction with Figure 2, the process can also be applied to other systems, and the present disclosure is not limited to this. In Figure 6, the EMS 220 performs model training to obtain the part of the bilateral AI model to be deployed to the terminal device 240 (referred to as the encoder model or UE-side model) and the part to be deployed to the access network device 230 (referred to as the decoder model or BS-side model).

[0153] Optionally, at 601 , the NMS 210 sends a device query request message to the EMS 220 ; and at 602 , the EMS 220 sends a device query response message to the NMS 210 .

[0154] Exemplarily, the device query request message can be used to request a query on the status information of the access network device and to request a query on the status information of the terminal device group. For example, the device query request message includes the identifier of the access network device 230 and the identifier of the terminal device group (such as including the terminal device 240). Accordingly, the device query response message is used to provide the status information of the queried access network device and the status information of the terminal device group. In some embodiments, the status information of the access network device may include one or more of the following: the computing power status, storage status, energy consumption status, and network energy consumption target value of the access network device. In some embodiments, the status information of the terminal device group may include one or more of the following: the computing power status, storage status, energy consumption status, and network energy consumption target value of each terminal device in the terminal device group.

[0155] At 603, the NMS 210 sends a fourth model training request to the EMS 220. The fourth model training request may further include one or more of the following: a model identifier of the bilateral AI model, an identifier of the access network device 230, an identifier of the terminal device group (including the terminal device 240), information indicating the type of inference task for which the bilateral AI model will be used, model performance requirements (including performance requirements of the decoder model and / or performance requirements of the encoder model), and a bilateral model training instruction.

[0156] In some embodiments, the fourth model training request may further include one or more of the following regarding the decoder model and / or encoder model: training energy consumption requirement, inference energy consumption requirement, model size requirement, or lightweight indication. For example, the NMS 210 may determine one or more of the following based on the device query response message: training energy consumption requirement, inference energy consumption requirement, model size requirement, or lightweight indication. For example, the NMS 210 may determine one or more of the following regarding the decoder model based on the status information of the access network device 230: a first training energy consumption requirement, a first inference energy consumption requirement, a first model size requirement, or a first lightweight indication. For example, the NMS 210 may determine one or more of the following regarding the encoder model based on the status information of the terminal device 240: a second training energy consumption requirement, a second inference energy consumption requirement, a second model size requirement, or a second lightweight indication.

[0157] Optionally, the fourth model training request may include an indication of requesting training data for UE-side model training, for example, the indication is "false", that is, no training data is required for the UE-side model.

[0158] At 604, the EMS 220 acquires training data. This process is similar to the embodiment described above in conjunction with 506 of FIG.

[0159] At 605, EMS 220 performs model training based on the fourth model training request using the training data to obtain a bilateral AI model, which includes a decoder model (i.e., a BS-side model) and an encoder model (i.e., a UE-side model). In this way, both parts of the bilateral AI model can be trained by EMS 220, avoiding the need for the server to participate in the training and eliminating the need to open additional training-related interfaces to the server, thereby reducing signaling interactions.

[0160] Optionally, at 606, EMS 220 may perform lightweight processing on the bilateral AI model. In some examples, if the fourth model training request includes a first lightweight indication and a second lightweight indication, indicating that lightweight processing is required, EMS 220 performs lightweight processing on the bilateral AI model (including the encoder model and the decoder model). In other examples, even if the fourth model training request does not include any lightweight indication, or the fourth model training request indicates that lightweight processing is not required, EMS 220 may decide at its own discretion to perform lightweight processing on the encoder model and / or decoder model in the bilateral AI model. The operation on lightweight processing is similar to the process described in 409 of Figure 4 above.

[0161] In this way, EMS 220 can obtain a trained bilateral AI model, including a trained encoder model and a trained decoder model, through 605, or through 605 and 606. It should be noted that although 606 is shown as a step independent of 605 in FIG6 , in other embodiments, 605 may include 606, that is, lightweight processing can be performed during the model training process.

[0162] At 607, the EMS 220 sends a fourth model training report to the NMS 210. Exemplarily, the fourth model training report includes a bilateral model indication to indicate that the bilateral AI model (including the encoder model and the decoder model) has been trained. Exemplarily, the fourth model training report may also include one or more of the following: a model identifier of the encoder model, an identifier of the terminal device group, a model identifier of the decoder model, an identifier of the access network device 230, the actual performance of the bilateral AI model, the actual training energy consumption of the bilateral AI model, the actual inference energy consumption of the bilateral AI model, the actual model size of the bilateral AI model, etc. For example, the actual performance of the bilateral AI model includes the actual performance of the encoder model and / or the actual performance of the decoder model, the actual training energy consumption of the bilateral AI model includes the actual training energy consumption of the encoder model and / or the actual training energy consumption of the decoder model, the actual inference energy consumption of the bilateral AI model includes the actual inference energy consumption of the encoder model and / or the actual inference energy consumption of the decoder model, the actual model size of the bilateral AI model includes the actual model size of the encoder model and / or the actual model size of the decoder model.

[0163] At 608, NMS 210 sends a model loading request to EMS 220. The model loading request is used to instruct EMS 220 to deploy the bilateral AI model. Specifically, the encoder model is deployed to the terminal device group (including terminal device 240) and the decoder model is deployed to the access network device 230. For example, the model loading request includes a bilateral model indication, a model identifier of the encoder model, an identifier of the terminal device group, a model identifier of the decoder model, and an identifier of the access network device 230.

[0164] At 609 , the EMS 220 loads the bilateral AI model to the access network device 230 based on the model loading request; and at 610 , the access network device 230 transfers the encoder model to the terminal device 240 .

[0165] Exemplarily, EMS 220 may send another model loading request to access network device 230 for deploying the encoder model to terminal device 240 and deploying the decoder model to access network device 230. In some examples, the decoder model and / or encoder model may be expressed in the form of model parameters or carried in the form of model files. For example, another model loading request includes the decoder model itself, or includes the model parameters of the decoder model or the file address of the model file of the decoder model. Another model loading request includes the encoder model itself, or includes the model parameters of the encoder model or the file address of the model file of the encoder model. Exemplarily, another model loading request may also include a bilateral model indication, a model identifier of the encoder model, an identifier of the terminal device group, a model identifier of the decoder model, and an identifier of access network device 230.

[0166] Illustratively, access network device 230 may send the encoder model to each terminal device in the group of terminal devices, including terminal device 240. In some examples, access network device 230 may send one or more of the following to terminal device 240: the encoder model, the bilateral model indication, a model identifier of the encoder model, and a model identifier of the decoder model.

[0167] At 611, the bilateral AI model can be used in a subsequent model reasoning process.

[0168] By combining the signaling interaction 600 described in FIG6 , the EMS 220 can complete the model training of the bilateral AI model, and it can be deployed at the terminal device 240 and the access network device 230, respectively. In some embodiments, during the training process, the status information of the access network device and the terminal device can be considered at the same time, so that the appropriate model can be trained based on the computing power, storage, and energy consumption requirements of the access network device and the terminal device. The fourth model training request and the fourth model training report may include a bilateral model indication, and the model loading request / another model loading request may include a bilateral model indication and / or bilateral model association information, thereby supporting the deployment and reasoning of the bilateral AI model.

[0169] In the above-mentioned embodiments in combination with Figures 4 to 6, model training is performed by the EMS 220 of the management plane, for example, a trained BS-side model (Figure 4 or Figure 5) or a trained bilateral AI model (Figure 6) is obtained. During the training process, the matching between the BS-side model and the UE-side model can be ensured through bilateral training instructions, that is, it is used as a bilateral AI model. When the model is deployed, the access network device and the terminal device can be enabled to determine the association between the BS-side model and the UE-side model through bilateral instructions, thereby avoiding considering the deployed model as a unilateral model. In this way, the matching and association between the BS-side model and the UE-side model during the model reasoning process can be ensured, thereby ensuring the performance of the model reasoning.

[0170] Figure 7 shows a schematic diagram of a signaling interaction 700 according to some embodiments of the present disclosure. Figure 7 involves a server 290, an NMS 210, an EMS 220, an access network device 230, and a terminal device 240. It can be understood that although the signaling interaction 700 is described in conjunction with Figure 2, the process can also be applied to other systems, and the present disclosure is not limited to this. In Figure 7, the server 290 and the NMS 210 perform model training. Specifically, the server 290 obtains the part of the bilateral AI model to be deployed to the terminal device 240 (referred to as the encoder model or UE-side model), and the NMS 210 obtains the part of the bilateral AI model to be deployed to the access network device 230 (referred to as the decoder model or BS-side model).

[0171] At 701, NMS 210 may open AI / ML management services to server 290. At 702, server 290 obtains training data. At 703, server 290 performs model training based on the training data to obtain an encoder model (i.e., a UE-side model) and an initial decoder model. At 704, server 290 sends a first model training request to NMS 210. Optionally, at 705, NMS 210 sends a device query request message to EMS 220; and at 706, EMS 220 sends a device query response message to NMS 210. Specifically, with respect to 701 to 706, reference may be made to 401 to 406 described in conjunction with FIG. 4 , respectively, and for the sake of brevity, they will not be repeated here.

[0172] At 707, the NMS 210 performs model training based on the first model training request. Specifically, the NMS 210 obtains the decoder model (i.e., the BS-side model) through training using the training data for training the decoder model (i.e., the BS-side model) indicated in the first model training request and the structural information of the decoder model.

[0173] Optionally, at 708, the NMS 210 performs lightweight processing on the decoder model. Specifically, the NMS 210 may determine one or more of the following regarding the decoder model based on the device query response message received at 706: training energy consumption requirement, inference energy consumption requirement, or model size requirement. For example, the NMS 210 may decide to perform lightweight processing based on model performance requirements, training energy consumption requirements, inference energy consumption requirements, model size requirements, etc. Exemplarily, the NMS 210 may perform quantization, pruning, and other processing on the decoder model to obtain a decoder model after lightweight processing. It can be understood that lightweight processing is to compress the model while minimizing the model performance requirements (such as accuracy) to reduce the model's computational complexity, size (such as the number of parameters), etc.

[0174] In this way, the NMS 210 can obtain a trained decoder model that matches the encoder model trained by the server 290 through 707, or through 707 and 708. It should be noted that although 708 is shown as a step independent of 707 in FIG7 , in other embodiments, 707 may include 708, that is, lightweight processing may be performed during the model training process.

[0175] At 709 , the NMS 210 sends a first model training report to the server 290 to indicate that the decoder model has been trained. For example, the first model training report may include a model identifier of the trained decoder model and actual performance of the decoder model.

[0176] At 710, server 290 sends the encoder model to terminal device 240. At 711, server 290 sends a model delivery notification to NMS 210, indicating that the encoder model has been deployed on terminal device 240. Specifically, steps 710 to 711 may refer to steps 412 to 413 described in conjunction with FIG. 4 , and for the sake of brevity, they are not repeated here.

[0177] At 712, the NMS 210 loads the decoder model to the access network device 230. For example, the NMS 210 may send a model loading request to the access network device 230. For example, the NMS 210 may deploy the decoder model to the access network device 230 via the EMS 220. For example, the NMS 210 may send the model loading request to the EMS 220, which then forwards the model loading request to the access network device 230. For example, the model loading request may include the decoder model itself, or model parameters of the decoder model, or the file address of a model file for the decoder model. For example, the model loading request may also include a bilateral model indication, a model identifier of the encoder model, and a terminal device group identifier.

[0178] At 713 , the bilateral AI model may be used in a subsequent model reasoning process.

[0179] By combining the signaling interaction 700 described in FIG7 , the server 290 can be supported to call the model management service, and then the server 290 and the NMS 210 can collaboratively complete the model training of the bilateral AI model, and are deployed at the terminal device 240 and the access network device 230, respectively. In some embodiments, during the training process, the status information of the access network device 230 can be considered at the same time, so that a suitable model can be trained based on the computing power, storage, and energy consumption requirements of the access network device 230. The service types (i.e., model training service, model loading service) opened by the management plane (such as NMS 210) can be increased to support the server 290 outside the 3GPP network domain 250 to call the model training service and model loading service of the management plane. The first model training request can include a bilateral model indication, a model identifier of the UE-side model, a model structure of the BS-side model, and training data of the BS-side model, thereby supporting the management plane (such as NMS 210) to train the BS-side model with the training data provided by the external server 290. The model loading request may include a bilateral model indication and / or bilateral model association information, thereby indicating that the UE-side model has been deployed, thereby supporting the deployment and reasoning of the bilateral AI model.

[0180] Figure 8 shows a schematic diagram of another signaling interaction 800 according to some embodiments of the present disclosure. Figure 8 involves a server 290, an NMS 210, an EMS 220, an access network device 230, and a terminal device 240. It is understandable that although the signaling interaction 800 is described in conjunction with Figure 2, the process can also be applied to other systems, and the present disclosure is not limited to this. In Figure 8, the server 290 and the NMS 210 perform model training. Specifically, the server 290 obtains the part of the bilateral AI model to be deployed to the terminal device 240 (referred to as the encoder model or UE-side model), and the NMS 210 obtains the part of the bilateral AI model to be deployed to the access network device 230 (referred to as the decoder model or BS-side model).

[0181] At 801, NMS 210 may open AI / ML management services to server 290. At 802, server 290 sends a third model training request to NMS 210. Optionally, at 803, NMS 210 sends a device query request message to EMS 220; and at 804, EMS 220 sends a device query response message to NMS 210. Specifically, steps 801 to 804 may refer to steps 501 to 504 described in conjunction with FIG. For the sake of brevity, they are not repeated here.

[0182] At 805, NMS 210 obtains training data. Specifically, NMS 210 may obtain training data collected by terminal device 240 via EMS 220 and access network device 230. Specifically, terminal device 240 may perform operations such as data collection, measurement, and calculation to obtain training data, and terminal device 240 may send the training data to access network device 230 via an uplink. Access network device 230 may then provide the training data to EMS 220 (e.g., based on an instruction or request from EMS 220), so that NMS 210 can obtain the training data from EMS 220. Assuming that this embodiment is intended to train a bilateral AI model for CSI feedback, the training data may include raw CSI measured by terminal device 240.

[0183] At 806 , the NMS 210 performs model training using the training data based on the third model training request to obtain a decoder model (ie, a BS-side model) and an initial encoder model (ie, a UE-side model).

[0184] Optionally, at 807, the NMS 210 may perform lightweight processing on the decoder model. Specifically, based on the device query response message received at 804, the NMS 210 may determine one or more of the following requirements regarding the decoder model: training energy consumption requirement, inference energy consumption requirement, or model size requirement. For example, the NMS 210 may decide to perform lightweight processing based on model performance requirements, training energy consumption requirement, inference energy consumption requirement, model size requirement, etc. Exemplarily, the NMS 210 may perform quantization, pruning, and other processes on the decoder model to obtain a lightweight decoder model. In this way, the NMS 210 may obtain a trained decoder model through 806, or through both 806 and 807. It should be noted that although 807 is shown as a separate step from 806 in Figure 8, in other embodiments, 806 may include 807, meaning that lightweight processing can be performed during model training.

[0185] At 808, the NMS 210 sends a third model training report to the server 290. For example, the third model training report may indicate that the decoder model has been trained, and the third model training report includes training data for training the encoder model (ie, UE-side model) and structural information of the encoder model.

[0186] Exemplarily, after completing the model training at 806, the NMS 210 may construct training data for training the encoder model (i.e., the UE-side model) by the server 290 based on the training data obtained at 805 and, optionally, based on the initial encoder model obtained by the model training. Exemplarily, the third model training report may directly include the training data for training the encoder model (i.e., the UE-side model). Exemplarily, the third model training report may include a storage address of the training data for training the encoder model (i.e., the UE-side model), so that the server 290 may download the training data for training the encoder model (i.e., the UE-side model) from the storage address via FTP.

[0187] Exemplarily, the third model training report may include a model identifier of the trained decoder model and an identifier of the access network device 230. Exemplarily, the third model training report may also include one or more of the following: actual performance of the decoder model, actual training energy consumption of the decoder model, actual inference energy consumption of the decoder model, and actual model size of the decoder model.

[0188] At 809, the server 290 performs model training based on the third model training report to obtain a trained encoder model. Specifically, the server 290 can construct a model based on the structural information of the encoder model, and perform model training based on the training data for training the encoder model (i.e., the UE-side model). It can be understood that since the training at 809 is based on the training data and structural information for training the encoder model provided by the NMS 210, the trained encoder model can match the decoder model trained by the NMS 210. In some examples, during the training process at 809, the model performance requirements for the encoder model in the third model training report can also be considered to ensure that the trained encoder model can adapt to the status information at the terminal device and ensure the availability of the encoder model.

[0189] At 810, server 290 sends the encoder model to terminal device 240. At 811, server 290 sends a model delivery notification to NMS 210, indicating that the encoder model has been deployed on terminal device 240. At 812, NMS 210 can deploy the decoder model to access network device 230 via EMS 220. At 813, the bilateral AI model can be used for subsequent model inference. Specifically, steps 810 through 813 can be referenced to steps 710 through 713 described in conjunction with FIG. For the sake of brevity, these steps are not repeated here.

[0190] By combining the signaling interaction 800 described in FIG8 , the server 290 can be supported to call the model management service, and then the server 290 and the NMS 210 can collaboratively complete the model training of the bilateral AI model, and are deployed at the terminal device 240 and the access network device 230, respectively. In some embodiments, during the training process, the status information of the access network device 230 can be considered at the same time, so that a suitable model can be trained based on the computing power, storage, and energy consumption requirements of the access network device 230. The service types (i.e., model training service, model loading service) opened by the management plane (such as NMS 210) can be increased to support the server 290 outside the 3GPP network domain 250 to call the model training service and model loading service of the management plane. The third model training request can include a bilateral model indication and an indication of training data for requesting UE-side model training, and the third model training report can include training data for training the encoder model and structural information of the encoder model, thereby supporting the server 290 outside the 3GPP network domain 250 to use the training data provided by the management plane to train the UE-side model. The model loading request may include a bilateral model indication and / or bilateral model association information, thereby indicating that the UE-side model has been deployed, thereby supporting the deployment and reasoning of the bilateral AI model.

[0191] Figure 9 shows a schematic diagram of another signaling interaction 900 according to some embodiments of the present disclosure. Figure 9 involves NMS 210, EMS 220, access network device 230 and terminal device 240. It can be understood that although the signaling interaction 900 is described in conjunction with Figure 2, the process can also be applied to other systems, and the present disclosure is not limited to this. In Figure 9, the NMS 210 performs model training to obtain the part of the bilateral AI model to be deployed to the terminal device 240 (called the encoder model or UE-side model) and the part to be deployed to the access network device 230 (called the decoder model or BS-side model).

[0192] Optionally, at 901, NMS 210 sends a device query request message to EMS 220; and at 902, EMS 220 sends a device query response message to NMS 210. Specifically, 901 and 902 may refer to 601 and 602 described in conjunction with FIG. 6 , and are not repeated here for brevity.

[0193] At 903, NMS 210 obtains training data. Specifically, NMS 210 may obtain training data collected by terminal device 240 via EMS 220 and access network device 230. This process may refer to 805 described above in conjunction with FIG. 8 and will not be repeated here.

[0194] At 904, NMS 210 uses the training data to perform model training to obtain a bilateral AI model, which includes a decoder model (i.e., a BS-side model) and an encoder model (i.e., a UE-side model). In this way, both parts of the bilateral AI model can be trained by NMS 210, eliminating the need for the server to participate in training and the need to open additional training-related interfaces to the server, thereby reducing signaling interactions.

[0195] Optionally, at 905, NMS 210 may perform lightweight processing on the bilateral AI model. Specifically, NMS 210 may determine one or more of the following regarding the decoder model and / or encoder model based on the device query response message received at 902: training energy consumption requirement, inference energy consumption requirement, or model size requirement. For example, NMS 210 may determine one or more of the following regarding the decoder model based on the status information of access network device 230: a first training energy consumption requirement, a first inference energy consumption requirement, or a first model size requirement. For example, NMS 210 may determine one or more of the following regarding the encoder model based on the status information of terminal device 240: a second training energy consumption requirement, a second inference energy consumption requirement, or a second model size requirement.

[0196] For example, NMS 210 may decide to perform lightweight processing on the encoder model and / or decoder model based on the first and / or second model performance requirements, the first and / or second training energy consumption requirements, the first and / or second inference energy consumption requirements, the first and / or second model size requirements, etc. Exemplarily, NMS 210 may perform quantization, pruning, and other processing on the encoder model and / or decoder model to obtain a bilateral AI model after lightweight processing. In this way, NMS 210 may obtain a trained bilateral AI model, including an encoder model and a decoder model, through 904, or through 904 and 905. It should be noted that although 905 is shown as a step independent of 904 in Figure 9, in other embodiments, 904 may include 905, that is, lightweight processing may be performed during model training.

[0197] At 906 , the NMS 210 may load the bilateral AI model to the access network device 230 via the EMS 220 , and at 907 , the access network device 230 may transfer the encoder model to the terminal device 240 . For example, the NMS 210 may send a model loading request to the access network device 230 .

[0198] For example, NMS 210 may send a model loading request to EMS 220, which then forwards the model loading request to access network device 230. For example, the model loading request may include the decoder model itself, or model parameters of the decoder model, or the file address of a model file for the decoder model; the model loading request may also include the encoder model itself, or model parameters of the encoder model, or the file address of a model file for the encoder model. For example, the model loading request may also include a bilateral model indication, a model identifier of the encoder model, a model identifier of the decoder model, an identifier of the access network device 230, and a terminal device group identifier.

[0199] Illustratively, access network device 230 may send the encoder model to each terminal device in the group of terminal devices, including terminal device 240. In some examples, access network device 230 may send one or more of the following to terminal device 240: the encoder model, the bilateral model indication, a model identifier of the encoder model, and a model identifier of the decoder model.

[0200] At 908 , the bilateral AI model may be used in a subsequent model inference process.

[0201] By combining the signaling interaction 900 described in conjunction with FIG9 , the NMS 210 can complete the model training of the bilateral AI model and deploy it on the terminal device 240 and the access network device 230, respectively. In some embodiments, during the training process, the status information of the access network device and the terminal device can be considered simultaneously, so that the appropriate model can be trained based on the computing power, storage, and energy consumption requirements of the access network device and the terminal device. The model loading request can include a bilateral model indication and / or bilateral model association information, thereby supporting the deployment and reasoning of the bilateral AI model.

[0202] In the above-mentioned embodiments in combination with Figures 7 to 9, model training is performed by the NMS 210 of the management plane, for example, a trained BS-side model (Figure 7 or Figure 8) or a trained bilateral AI model (Figure 9) is obtained. During the training process, the matching between the BS-side model and the UE-side model can be ensured through bilateral training instructions, that is, it is used as a bilateral AI model. When the model is deployed, the access network device and the terminal device can be enabled to determine the association between the BS-side model and the UE-side model through bilateral instructions, thereby avoiding considering the deployed model as a unilateral model. In this way, the matching and association between the BS-side model and the UE-side model during the model reasoning process can be ensured, thereby ensuring the performance of the model reasoning.

[0203] Figure 10A illustrates a training scenario 1010 according to some embodiments of the present disclosure. Scenario 1010 corresponds to an embodiment in which a server trains a UE-side model and an NMS or EMS trains a BS-side model, as shown in Figures 4 or 7 above. As shown in Figure 10A , the server acquires training data from the UE and provides the training data for training the BS-side model to the NMS or EMS. Furthermore, the server deploys the trained UE-side model to the UE, while the NMS or EMS deploys the trained BS-side model to the BS.

[0204] Figure 10B illustrates a schematic diagram of a training scenario 1020 according to some embodiments of the present disclosure. Scenario 1020 corresponds to an embodiment in which a server trains a UE-side model and an NMS or EMS trains a BS-side model, as shown in Figure 5 or Figure 8 above. As shown in Figure 10B , the NMS or EMS obtains training data from the UE via the BS and provides the training data for training the UE-side model to the server. Furthermore, the server deploys the trained UE-side model to the UE, while the NMS or EMS deploys the trained BS-side model to the BS.

[0205] Figure 10C illustrates a training scenario 1030 according to some embodiments of the present disclosure. Scenario 1030 corresponds to an embodiment in which an NMS or EMS trains both a UE-side model and a BS-side model, as described above in Figures 6 or 9 . As shown in Figure 10C , the NMS or EMS obtains training data from the UE via the BS. Furthermore, the NMS or EMS provides both the trained BS-side model and the trained UE-side model to the BS, which then deploys the UE-side model to the UE.

[0206] Through the example embodiments of the present disclosure as described above, the model training interface for bilateral AI model training is enhanced, such as the interface between the server 290 and the NMS 210, and the interface between the NMS 210 and the EMS 220, so that the collaboration between the server 290 and the management plane (including the NMS 210 and the EMS 220) can be enhanced; the model loading interface for bilateral AI model loading is also enhanced, such as the interface between the server 290 and the NMS 210, between the NMS 210 and the EMS 220, and between the NMS 210 / EMS 220 and the access network device 230, and between the server 290 and the terminal device 240, so that the successful deployment of the model can be determined.

[0207] In this way, the embodiments of the present disclosure implement the management plane to assist the training process of the bilateral AI model over the air interface, thereby supporting the use (inference) of the bilateral AI model over the air interface.

[0208] It should be understood that the embodiments of Figures 4 to 10 above are described in conjunction with the system 200 shown in Figure 2, but the embodiments of the present disclosure may also be applied to the system 300 shown in Figure 3 or other systems not shown herein. The embodiments of Figures 4 to 10 above involve interactions between different devices (or network elements, or systems), but the embodiments of the present disclosure do not limit the specific form of the interactive signaling, which may be in various forms such as messages, notifications, commands, requests, responses, reports, information, and signaling.

[0209] It should be understood that although the encoder model and decoder model are used to represent the UE-side model and the BS-side model, respectively, in actual situations, the bilateral AI model is not limited to the encoder and decoder architecture. For example, the bilateral AI model can include a first part deployed at the access network device and a second part deployed at the terminal device. For example, the encoder model and decoder model in the above embodiment can be replaced by the second part and the first part, respectively.

[0210] It should also be understood that the division of the modes, situations, categories and embodiments in the embodiments of the present disclosure is only for the convenience of description and should not constitute a special limitation. The features in various modes, categories, situations and embodiments can be combined with each other when it is logical.

[0211] It should also be understood that the above content is only intended to help those skilled in the art better understand the embodiments of the present disclosure, and is not intended to limit the scope of the embodiments of the present disclosure. Those skilled in the art may make various modifications, variations, or combinations based on the above content. Such modifications, variations, or combinations are also within the scope of the embodiments of the present disclosure.

[0212] It should also be understood that the description of the above content focuses on emphasizing the differences between the various embodiments, and the same or similar points can be referenced or borrowed from each other. For the sake of brevity, they will not be repeated here.

[0213] FIG11 shows a schematic block diagram of a network management device 1100 according to some embodiments of the present disclosure. The network management device 1100 can be implemented as the NMS 210 or EMS 220 shown in FIG2 , or as a part of the NMS 210 or EMS 220 (such as a chip), etc., and the present disclosure is not limited to this. The network management device 1100 can be implemented as the non-RT RIC 310 or near-RT RIC 320 shown in FIG3 , or as a part of the non-RT RIC 310 or near-RT RIC 320 (such as a chip), etc., and the present disclosure is not limited to this. As shown in FIG11 , the network management device 1100 includes a training data acquisition module 1110, a model training module 1120, and a model sending module 1130.

[0214] The training data acquisition module 1110 is configured to acquire training data for training the first part of a bilateral model, where the bilateral model includes a first part to be deployed on a network device and a second part to be deployed on a terminal device. The model training module 1120 is configured to perform model training based on the training data to obtain the first part. The model deployment module 1130 is configured to send the first part to the network device. For example, the network device may be the access network device 230 shown in FIG. 2 or the access network device 330 shown in FIG. 3 .

[0215] In some embodiments, the model training module 1120 may be further configured to perform lightweight processing on the first part; or the network management device 1100 may further include a lightweight processing module configured to perform lightweight processing on the first part. Exemplarily, the lightweight first part meets at least one of the following: a performance requirement for the first part, an energy consumption requirement for training the first part, an energy consumption requirement for an inference task for which the first part will be used, or a first model size requirement for the first part.

[0216] In some implementations of the present disclosure, the network management device 1100 may be implemented as a second network management device, such as the EMS 220 or the near-RT RIC 320 .

[0217] In some examples, the training data acquisition module 1110 is specifically configured to receive a first training request from a first network management device, the first training request indicating training a first part of the bilateral model, the first training request including training data for training the first part of the bilateral model and structural information of the first part.

[0218] In some examples, the network management device 1100 may also include a reporting module configured to send a first model training report to the first network management device, the first model training report including at least one of the following: a model identification of the bilateral model, a model identification of the first part, actual performance of the first part, actual energy consumption of training the first part, estimated energy consumption of the inference task for which the first part will be used, or an actual model size of the first part.

[0219] In some examples, the network management device 1100 may further include a receiving module configured to receive a model loading request from the first network management device, the model loading request instructing to load the first part to the network device. The model sending module 1130 may be configured to send the first part to the network device based on the model loading request.

[0220] Optionally, the model loading request may include at least one of the following: a first identifier of the network device, a model identifier of the first part, a second identifier of the terminal device, a model identifier of the second part, or a model identifier of the bilateral model.

[0221] Optionally, the first training request also includes at least one of the following: a model identifier of the bilateral model, a model identifier of the first part, a model identifier of the second part, a first identifier of the network device, a second identifier of the terminal device, indication information for indicating the type of reasoning task for which the bilateral model will be used, performance requirements of the bilateral model, energy consumption requirements for bilateral model training, energy consumption requirements for reasoning tasks, a first model size requirement for the first part, or a lightweight indication.

[0222] In some examples, the network management device 1100 may further include a request receiving module and a response sending module. The request receiving module is configured to receive a request message from the first network management device, the request message being used to request status information of the network device. The response sending module is configured to send a response message to the first network management device in response to the request message, the response message including at least one of the following: computing power status of the network device, storage status of the network device, energy consumption information of the network device, or a network energy consumption target value of the network device.

[0223] In some examples, the network management device 1100 may further include a reporting module configured to send a second model training report to the first network management device, where the second model training report includes training data for training the second part and structural information of the second part.

[0224] Optionally, the second model training report also includes at least one of the following: the model identification of the bilateral model, the model identification of the first part, the first identification of the network device, the actual performance of the first part, the actual energy consumption of the training of the first part, the estimated energy consumption of the inference task for which the first part will be used, the actual model size of the first part, the performance requirement of the second part, the second energy consumption requirement of the inference task for which the second part will be used, or the second model size requirement of the second part.

[0225] The network management device 1100 is implemented as a second network management device, and can be used to implement the various processes performed by the EMS 220 in conjunction with Figures 4 to 6 above, which will not be described here for brevity.

[0226] In some other implementations of the present disclosure, the network management device 1100 may be implemented as a first network management device, such as the NMS 210 or the non-RT RIC 310 .

[0227] In some examples, the training data acquisition module 1110 is specifically configured to receive a second training request from the server, the second training request indicating training the first part of the bilateral model, and the second training request including training data for training the first part of the bilateral model and structural information of the first part.

[0228] In some examples, the network management device 1100 may also include a reporting module configured to send a fourth model training report to the server, the fourth model training report including at least one of the following: a model identification of the bilateral model, a model identification of the first part, or actual performance of the first part.

[0229] In some examples, the network management device 1100 may further include a receiving module configured to receive a notification message from the server, the notification message indicating that the second part trained by the server has been loaded into the terminal device. The model sending module 1130 may be configured to send the first part to the network device in response to the notification message.

[0230] Optionally, the second training request also includes at least one of the following: a model identifier of the bilateral model, a model identifier of the first part, a model identifier of the second part, a first identifier of the network device, a second identifier of the terminal device, indication information for indicating the type of inference task for which the bilateral model will be used, or performance requirements of the bilateral model.

[0231] In some examples, the network management device 1100 may further include a request sending module and a response receiving module. The request sending module is configured to send a request message to the second network management device, the request message including the first identifier of the network device. The response receiving module is configured to receive a response message from the second network management device, the response message including at least one of the following: computing power status of the network device, storage status of the network device, energy consumption information of the network device, or a network energy consumption target value of the network device.

[0232] Exemplarily, the network management device 1100 may also include a determination module configured to determine at least one of the following based on the response message: the energy consumption requirement of bilateral model training, the first energy consumption requirement of the inference task to be used for the first part, the second energy consumption requirement of the inference task to be used for the second part, the first model size requirement of the first part, or the second model size requirement of the second part.

[0233] In some examples, the network management device 1100 may further include a reporting module configured to send a second model training report to the server, where the second model training report includes training data for training the second part and structural information of the second part.

[0234] Optionally, the second model training report also includes at least one of the following: the model identification of the bilateral model, the model identification of the first part, the first identification of the network device, the actual performance of the first part, the actual energy consumption of the training of the first part, the estimated energy consumption of the inference task for which the first part will be used, the actual model size of the first part, the performance requirement of the second part, the second energy consumption requirement of the inference task for which the second part will be used, or the second model size requirement of the second part.

[0235] The network management device 1100 is implemented as a first network management device, and can be used to implement the various processes performed by the NMS 210 in conjunction with Figures 7 to 9 above, which will not be described again for brevity.

[0236] Figure 12 shows a schematic block diagram of a network management device 1200 according to some embodiments of the present disclosure. Network management device 1200 can be implemented as NMS 210 as shown in Figure 2 , or as a portion of NMS 210 (e.g., a chip), etc., without limitation in this disclosure. Network management device 1100 can be implemented as non-RT RIC 310 as shown in Figure 3 , or as a portion of non-RT RIC 310 (e.g., a chip), etc., without limitation in this disclosure. As shown in Figure 12 , network management device 1200 includes a sending module 1210 and a receiving module 1220.

[0237] The sending module 1210 is configured to send a first training request to the second network management device, where the first training request indicates training of a first part of a bilateral model, where the bilateral model includes a first part to be deployed at the network device and a second part to be deployed at the terminal device. The receiving module 1220 is configured to receive a model training report from the second network management device, where the model training report includes at least a model identifier of the first part.

[0238] Optionally, the first training request includes: training data for training the first part, and structural information of the first part.

[0239] Optionally, the model training report includes a first model training report, which includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, an actual performance of the first part, an actual energy consumption of training the first part, an estimated energy consumption of an inference task for which the first part will be used, or an actual model size of the first part.

[0240] In some examples, the receiving module 1220 is further configured to receive a second training request from the server, the second training request indicating training the first part of the bilateral model, the second training request including: training data for training the first part, and structural information of the first part.

[0241] Optionally, the model training report includes a second model training report, and the second model training report includes training data for training the second part and structural information of the second part.

[0242] Optionally, the second model training report also includes at least one of the following: the model identification of the bilateral model, the model identification of the first part, the first identification of the network device, the actual performance of the first part, the actual energy consumption of the training of the first part, the estimated energy consumption of the inference task for which the first part will be used, and the actual model size of the first part.

[0243] In some examples, the sending module of the network management device 1200 may be further configured to forward the second model training report to the server, and the receiving module may be further configured to receive a notification message from the server indicating that the second part has been loaded into the terminal device. The sending module may also be configured to send a model loading request to the second network management device, the model loading request indicating that the first part is loaded into the network device.

[0244] The network management device 1200 is implemented as a first network management device, and can be used to implement the various processes performed by the NMS 210 in conjunction with Figures 4 to 6 above, which will not be described in detail here for the sake of brevity.

[0245] Figure 13 shows a schematic block diagram of a server 1300 according to some embodiments of the present disclosure. The server 1300 can be implemented as the server 290 shown in Figure 2, or as a part of the server 290 (such as a chip), etc., which is not limited by the present disclosure. The server 1300 can be implemented as the server 390 shown in Figure 3, or as a part of the server 390 (such as a chip), etc., which is not limited by the present disclosure. As shown in Figure 13, the server 1300 includes a training data acquisition module 1310, a model training module 1320, and a model sending module 1330.

[0246] The training data acquisition module 1310 is configured to acquire training data for training the second part of the bilateral model, where the bilateral model includes a first part to be deployed on the network device and a second part to be deployed on the terminal device. The model training module 1320 is configured to perform model training based on the training data to obtain the second part. The model transmission module 1330 is configured to transmit the second part to the terminal device.

[0247] In some examples, server 1300 also includes a sending module configured to send a second training request to the first network management device, the second training request indicating training the first part of the bilateral model, the second training request including training data for training the first part of the bilateral model, and structural information of the first part.

[0248] In some examples, server 1300 further includes a receiving module configured to receive a first model training report from a first network management device, where the first model training report indicates that the first part has been trained.

[0249] In some examples, the training data acquisition module 1310 is specifically configured to receive a second model training report from the first network management device, where the second model training report includes training data for training the second part and structural information of the second part.

[0250] Optionally, the second model training report also includes at least one of the following: the model identification of the bilateral model, the model identification of the first part, the first identification of the network device, the actual performance of the first part, the performance requirement of the second part, the second energy consumption requirement of the inference task for which the second part will be used, or the second model size requirement of the second part.

[0251] In some examples, the server 1300 further includes a sending module configured to send a notification message to the first network management device, where the notification message indicates that the second part has been loaded into the terminal device.

[0252] The server 1300 can be used to implement the various processes performed by the server 290 in conjunction with Figures 4 to 5 and Figures 7 to 8, which are not described here for brevity.

[0253] The division of modules or units in the embodiments of the present disclosure is illustrative and is merely a logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the disclosed embodiments may be integrated into a single unit, exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0254] Embodiments of the present disclosure also provide a system that may include a first network management device, a second network management device, a server, a network device, and a terminal device. Exemplarily, the server may be server 290 or server 390, which may obtain the second part of the bilateral model through model training. Exemplarily, the first network management device may be NMS 210 or non-RT RIC 310, which may receive a second training request from the server. The first network management device may also send a first training request to the second network management device, or the first network management device may obtain the first part of the bilateral model through training. Exemplarily, the second network management device may be EMS 220 or near-RT RIC 320, which may receive the first training request from the first network management device and perform model training based on the first training request to obtain the first part. Exemplarily, the network device may be access network device 230 or access network device 330 (such as O-CU 332 and O-DU 334), which may receive the first part from the first network management device or the second network management device. Optionally, the network device may also receive the second part and send the second part to the terminal device. Exemplarily, the terminal device may be the terminal device 240 or 340, which may receive the second part from the network device. The second part may be obtained through training by the server, the first network management device, or the second network management device.

[0255] FIG14 shows a schematic block diagram of an example device 1400 that can be used to implement an embodiment of the present disclosure. Device 1400 can be implemented as or included in the aforementioned first network management device, second network management device, or server. For example, device 1400 can be implemented as NMS 210, EMS 220, or server 290 in FIG2 , or non-RT RIC 310, near-RT RIC 320, or server 390 in FIG3 .

[0256] As shown in FIG. 14 , the device 1400 includes one or more processors 1410 , one or more memories 1420 coupled to the processors 1410 , and a communication module 1440 coupled to the processors 1410 .

[0257] The communication module 1440 can be used for two-way communication. The communication module 1440 can have at least one communication interface for communication. The communication interface can include any interface necessary for communicating with other devices.

[0258] Processor 1410 can be any type suitable for the local technology network and can include, but is not limited to, at least one of the following: a general-purpose computer, a special-purpose computer, a microcontroller, a digital signal processor (DSP), or one or more of a controller-based multi-core controller architecture. Device 1400 can have multiple processors, such as application-specific integrated circuit chips, which are time-slave to a clock synchronized with a main processor.

[0259] The memory 1420 may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, at least one of the following: read-only memory (ROM) 1424, erasable programmable read-only memory (EPROM), flash memory, hard disk, compact disc (CD), digital video disc (DVD), or other magnetic and / or optical storage. Examples of volatile memories include, but are not limited to, at least one of the following: random access memory (RAM) 1422, or other volatile memories that do not persist during a power outage.

[0260] Computer program 1430 includes computer executable instructions executed by associated processor 1410. Program 1430 may be stored in ROM 1424. Processor 1410 may perform any suitable actions and processes by loading program 1430 into RAM 1422.

[0261] The embodiments of the present disclosure may be implemented with the aid of program 1430 so that device 1400 may perform any of the processes discussed with reference to Figures 4 to 10. The embodiments of the present disclosure may also be implemented by hardware or by a combination of software and hardware.

[0262] The program 1430 may be tangibly embodied in a computer-readable medium that may be included in the device 1400 (such as in the memory 1420) or other storage device accessible by the device 1400. The program 1430 may be loaded from the computer-readable medium into the RAM 1422 for execution. The computer-readable medium may include any type of tangible, non-volatile memory, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc.

[0263] In some embodiments, the communication module 1440 in the device 1400 may be implemented as a transmitter and a receiver (or transceiver), which may be configured to send / receive training requests, training reports, notifications, loading requests, etc. In addition, the device 1400 may further include one or more of a scheduler, a controller, and a radio frequency / antenna, which will not be elaborated in detail in this disclosure.

[0264] For example, the device 1400 in FIG. 14 may be implemented as an electronic device, or may be implemented as a chip or a chip system in an electronic device, which is not limited in the embodiments of the present disclosure.

[0265] The present disclosure also provides a chip that may include an input interface, an output interface, and a processing circuit. In the present disclosure, the input interface and the output interface may be used to implement signaling or data interaction, while the processing circuit may be used to implement signaling or data information generation and processing.

[0266] The embodiments of the present disclosure further provide a chip system, including a processor for supporting a computing device to implement the functions involved in any of the above embodiments. In one possible design, the chip system may also include a memory for storing necessary program instructions and data. When the processor executes the program instructions, the device in which the chip system is installed implements the method involved in any of the above embodiments. Exemplarily, the chip system may be composed of one or more chips, or may include chips and other discrete devices.

[0267] An embodiment of the present disclosure further provides a processor for coupling with a memory, wherein the memory stores instructions. When the processor executes the instructions, the processor executes the methods and functions involved in any of the above embodiments.

[0268] The embodiments of the present disclosure further provide a computer program or computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the methods and functions involved in any of the above embodiments.

[0269] An embodiment of the present disclosure further provides a non-transitory computer-readable storage medium having computer instructions stored thereon. When a processor executes the instructions, the processor executes the methods and functions involved in any of the above embodiments.

[0270] In general, various embodiments of the present application can be implemented in hardware or dedicated circuits, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while other aspects can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of the present disclosure are shown and described as block diagrams, flow charts, or using some other graphical representation, it should be understood that the blocks, devices, systems, techniques, or methods described herein can be implemented as, by way of non-limiting example, hardware, software, firmware, dedicated circuits or logic, general-purpose hardware or a controller or other computing device, or some combination thereof.

[0271] The present disclosure also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which are executed in a device on a real or virtual processor of a target to perform the process / method as described above with reference to the accompanying drawings. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided between program modules as needed. The machine-executable instructions for the program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0272] The computer program code for implementing the disclosed method can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device so that the program code, when executed by the computer or other programmable data processing device, causes the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or entirely on a remote computer or server.

[0273] In the context of the present disclosure, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic, or other forms of propagated signals, such as carrier waves, infrared signals, and the like.

[0274] A computer-readable medium may be any tangible medium that contains or stores a program for or in connection with an instruction execution system, apparatus, or device. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination thereof. More detailed examples of computer-readable storage media include an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0275] In addition, although the operations of the method of the present disclosure are described in a particular order in the accompanying drawings, this does not require or imply that these operations must be performed in this particular order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can change the order of execution. Additionally or alternatively, certain steps can be omitted, multiple steps can be combined into one step, and / or one step can be decomposed into multiple steps. It should also be noted that the features and functions of two or more devices according to the present disclosure can be embodied in one device. Conversely, the features and functions of a device described above can be further divided into being embodied by multiple devices.

[0276] While various implementations of the present disclosure have been described above, the above descriptions are exemplary, non-exhaustive, and not intended to limit the disclosed implementations. Many modifications and variations will be apparent to those skilled in the art without departing from the scope of the described implementations. The terminology used herein is selected to provide a good explanation of the principles, practical applications, or improvements to existing technologies, or to enable other persons skilled in the art to understand the various implementations disclosed herein.

Claims

1. A method for model training, comprising: Acquire training data for training a first part of a bilateral model, wherein the bilateral model includes the first part to be deployed at a network device and a second part to be deployed at a terminal device; Performing model training based on the training data to obtain the first part; as well as The first portion is sent to the network device.

2. The method according to claim 1, further comprising: The first part is lightweighted so that the lightweighted first part meets at least one of the following: a performance requirement of the first part, an energy consumption requirement of training the first part, an energy consumption requirement of an inference task for which the first part will be used, or a first model size requirement of the first part.

3. The method according to claim 1 or 2, wherein: The obtaining of the first part of training data for training the bilateral model includes: A first training request is received from a first network management device, wherein the first training request indicates training the first part of the bilateral model, and the first training request includes training data for training the first part of the bilateral model and structural information of the first part.

4. The method according to claim 3, further comprising: Send a first model training report to the first network management device, wherein the first model training report includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, an actual performance of the first part, an actual energy consumption of training the first part, an estimated energy consumption of an inference task for which the first part will be used, or an actual model size of the first part.

5. The method according to claim 3 or 4, further comprising: receiving a model loading request from the first network management device, the model loading request instructing to load the first part into the network device; as well as Based on the model loading request, the first portion is sent to the network device.

6. The method according to claim 5, wherein the model loading request includes at least one of the following: a first identifier of the network device, a model identifier of the first part, a second identifier of the terminal device, a model identifier of the second part, or a model identifier of the bilateral model.

7. The method according to any one of claims 3 to 6, wherein the first training request further comprises at least one of the following: the model identifier of the bilateral model, The model identification of the first part, The model identification of the second part, a first identifier of the network device, a second identifier of the terminal device, Indicative information for indicating the type of reasoning task for which the bilateral model is to be used, The performance requirements of the bilateral model, The energy consumption requirement of the bilateral model training, The energy consumption requirement of the inference task, a first model size requirement of the first part, or Lightweight instructions.

8. The method according to any one of claims 3 to 7, further comprising: receiving a request message from the first network management device, wherein the request message is used to request status information of the network device; as well as In response to the request message, a response message is sent to the first network management device, the response message including at least one of the following: the computing power status of the network device, the storage status of the network device, the energy consumption information of the network device, or the network energy consumption target value of the network device.

9. The method according to claim 1 or 2, wherein: The obtaining of the first part of training data for training the bilateral model includes: A second training request is received from a server, wherein the second training request indicates training the first part of the bilateral model, and the second training request includes the training data for training the first part of the bilateral model and structural information of the first part.

10. The method according to claim 9, further comprising: Sending a fourth model training report to the server, wherein the fourth model training report includes at least one of the following: a model of the bilateral model identification, a model identification of the first part, or an actual performance of the first part.

11. The method according to claim 9 or 10, further comprising: receiving a notification message from the server, wherein the notification message indicates that the second part obtained by training the server has been loaded into the terminal device; as well as In response to the notification message, the first portion is sent to the network device.

12. The method according to any one of claims 9 to 11, wherein the second training request further comprises at least one of the following: the model identifier of the bilateral model, The model identification of the first part, The model identification of the second part, a first identifier of the network device, a second identifier of the terminal device, Indicative information for indicating the type of reasoning task for which the bilateral model is to be used, or Performance requirements of the bilateral model.

13. The method according to any one of claims 9 to 12, further comprising: Sending a request message to the second network management device, wherein the request message includes a first identifier of the network device; receiving a response message from the second network management device, the response message including at least one of the following: a computing power status of the network device, a storage status of the network device, energy consumption information of the network device, or a network energy consumption target value of the network device; and Based on the response message, determine at least one of the following: the energy consumption requirement of the bilateral model training, the first energy consumption requirement of the reasoning task for which the first part will be used, the second energy consumption requirement of the reasoning task for which the second part will be used, the first model size requirement of the first part, or the second model size requirement of the second part.

14. The method according to claim 1 or 2, further comprising: A second model training report is sent to the first network management device or server, wherein the second model training report includes training data for training the second part and structural information of the second part.

15. The method according to claim 14, wherein the second model training report further includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, a first identification of the network device, an actual performance of the first part, an actual energy consumption of training the first part, an estimated energy consumption of the reasoning task for which the first part will be used, an actual model size of the first part, a performance requirement of the second part, a second energy consumption requirement of the reasoning task for which the second part will be used, or a second model size requirement of the second part.

16. A method for model training, comprising: Sending a first training request to a second network management device, the first training request indicating training a first part of a bilateral model, the bilateral model comprising the first part to be deployed at a network device and a second part to be deployed at a terminal device; as well as A model training report is received from the second network management device, wherein the model training report includes at least a model identifier of the first part.

17. The method of claim 16, wherein the first training request comprises: Training data for training the first part, and structural information of the first part.

18. A method according to claim 16 or 17, wherein the model training report includes a first model training report, and the first model training report includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, an actual performance of the first part, an actual energy consumption of training the first part, an estimated energy consumption of an inference task for which the first part will be used, or an actual model size of the first part.

19. The method according to claim 17 or 18, further comprising: A second training request is received from a server, where the second training request indicates training the first part of the bilateral model, and the second training request includes: training data for training the first part, and structural information of the first part.

20. The method according to claim 16, wherein the model training report comprises a second model training report, and the second model training report comprises training data for training the second part and structural information of the second part.

21. The method according to claim 20, wherein the second model training report further includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, a first identification of the network device, an actual performance of the first part, an actual energy consumption of training the first part, an estimated energy consumption of an inference task for which the first part will be used, and an actual model size of the first part.

22. The method according to claim 20 or 21, further comprising: Forwarding the second model training report to the server; receiving a notification message from the server, the notification message indicating that the second part has been loaded into the terminal device; as well as A model loading request is sent to the second network management device, where the model loading request indicates that the first part is loaded into the network device.

23. A method for model training, comprising: Acquire training data for training a second part of the bilateral model, wherein the bilateral model includes a first part to be deployed at a network device and the second part to be deployed at a terminal device; Performing model training based on the training data to obtain the second part; as well as The second part is sent to the terminal device.

24. The method according to claim 23, further comprising: A second training request is sent to the first network management device, where the second training request indicates training the first part of the bilateral model, and the second training request includes the training data for training the first part of the bilateral model and structural information of the first part.

25. The method according to claim 24, further comprising: A first model training report is received from a first network management device, wherein the first model training report indicates that the first part has been trained.

26. The method of claim 23, wherein obtaining training data for the second part of the bilateral model comprises: A second model training report is received from the first network management device, wherein the second model training report includes training data for training the second part and structural information of the second part.

27. The method according to claim 26, wherein the second model training report further includes at least one of the following: a model identification of the bilateral model, a model identification of the first part, a first identification of the network device, an actual performance of the first part, a performance requirement of the second part, a second energy consumption requirement of the inference task for which the second part will be used, or a second model size requirement of the second part.

28. The method according to any one of claims 23 to 27, further comprising: A notification message is sent to the first network management device, the notification message indicating that the second part has been loaded into the terminal device.

29. A method for model training, comprising: The method according to any one of claims 16 to 22 is performed by a first network management device, and The method according to any one of claims 1 to 8 is performed by a second network management device.

30. The method of claim 29, further comprising: The method according to any one of claims 23 to 28 is performed by a server.

31. The method according to claim 29 or 30, further comprising: The network device receives a first part of the bilateral model trained by the second network management device; The network device receives a second portion of the bilateral model from a server; The network device sends the second part to the terminal device; as well as The terminal device receives the second part from the network device.

32. A method for model training, comprising: The method according to any one of claims 1 to 2 and 9 to 15 is performed by a first network management device; as well as The method according to any one of claims 23 to 28 is performed by a server.

33. The method of claim 32, further comprising: The network device receives a first part of the bilateral model trained by the first network management device; The network device receives a second portion of the bilateral model from the server; The network device sends the second part to the terminal device; as well as The terminal device receives the second part from the network device.

34. A communication device, comprising a module or component for executing the method according to any one of claims 1 to 28.

35. A system comprising: A first network management device, configured to perform the method according to any one of claims 16 to 22; as well as The second network management device is configured to execute the method according to any one of claims 1-8.

36. The system of claim 35, further comprising: The server is configured to execute the method according to any one of claims 23 to 28.

37. The system of claim 35 or 36, further comprising: Network devices are configured as: Receiving a first part of the bilateral model trained by the second network management device; receiving a second portion of the bilateral model from a server; sending the second part to a terminal device; as well as The terminal device is configured to receive the second part from the network device.

38. A system comprising: A first network management device, configured to perform the method as claimed in any one of claims 1 to 2 and 9 to 15; as well as The server is configured to execute the method according to any one of claims 23 to 28.

39. The system of claim 38, further comprising: Network devices are configured as: Receiving a first part of the bilateral model trained by the first network management device; receiving a second portion of the bilateral model from a server; sending the second part to the terminal device; as well as The terminal device is configured to receive the second part from the network device.

Images