Model inference time and state determination in wireless communications

Abstract

This application generally relates to wireless communication, comprising a user equipment transmitting nominal inference times for M models and a base station receiving nominal inference times for the M models, where M is an integer of 2 or greater. The base station activates a subset of models from the M models, wherein the subset comprises N models activated by the base station, and N is less than M. The base station transmits and the user equipment receives an indication of the activated model subset. This application also generally relates to wireless communication, comprising a user equipment that determines one of a plurality of states of a model; and based on the state determines whether to perform at least one operation associated with the model, wherein the at least one operation includes model inference and / or model monitoring.

Description

Technical Field

[0001] This application generally relates to model inference time and state determination in wireless communication. Background Technology

[0002] In some wireless communication systems, base stations and / or user equipment can leverage artificial intelligence (AI) and / or machine learning (ML) to improve performance. For example, a wireless communication system may employ one or more AI / ML models to perform one or more functions or operations. However, performing operations using AI / ML models consumes computational power. In various situations, the computational power available to a device to perform a function using a model at a given point in time may be insufficient. Therefore, methods for coordinating computational power within a wireless communication system may be desirable. Summary of the Invention

[0003] This application relates to methods, systems, apparatuses, and devices for wireless communication. In some embodiments, a wireless communication method includes: transmitting nominal inference times of M models to a base station by a user equipment, wherein M is an integer of 2 or greater; receiving an indication of a subset of models from the M models activated by the base station by the user equipment, wherein the subset includes N models activated by the base station, wherein N is an integer of 1 or greater and N is less than M; and determining the actual inference time of the subset by the user equipment based on the activated model subset and the nominal inference times.

[0004] In some other embodiments, a wireless communication method includes: receiving, by a base station, nominal inference times of M models from a user equipment, wherein M is an integer of 2 or greater; activating a subset of models from the M models by the base station, wherein the subset includes N models activated by the base station, wherein N is an integer of 1 or greater and N is less than M; and transmitting an indication of the activated subset of models to the user equipment by the base station.

[0005] In some other embodiments, a wireless communication method includes: a user equipment determining one of a plurality of states of a model; and the user equipment determining, based on the state, whether to perform at least one operation associated with the model, wherein the at least one operation includes at least one of: model inference or model monitoring.

[0006] In some other embodiments, an apparatus, such as a network device, is disclosed. This apparatus may include one or more processors and one or more memories, wherein the one or more processors are configured to read computer code from the one or more memories to implement any of the methods described above.

[0007] In some other embodiments, a computer program product is disclosed. This computer program product may include a non-transitory computer-readable program medium having stored computer code thereon, which, when executed by one or more processors, causes the one or more processors to implement any of the methods described above.

[0008] The foregoing and other aspects and their embodiments are described in more detail in the accompanying drawings, description and claims. Attached Figure Description

[0009] Figure 1 A block diagram of an example wireless communication system is shown.

[0010] Figure 2 A flowchart of a wireless communication method is shown.

[0011] Figure 3 A flowchart of another wireless communication method is shown.

[0012] Figure 4 A flowchart of another wireless communication method is shown.

[0013] Figure 5 A timing diagram of an example implementation of the model pattern is shown.

[0014] Figure 6 A schematic diagram of the overall state transition of the model is shown. Detailed Implementation

[0015] The example headings in the following sections are provided to facilitate understanding of the disclosed subject matter and do not limit the scope of the claimed subject matter in any way. Therefore, one or more features of one example section may be combined with one or more features of another example section. Furthermore, the term "5G" is used for clarity, but the technology disclosed in this application is not limited to 5G technology and can also be used in wireless systems implementing other protocols, such as 5G-evolution (5G-A), 6G, or higher.

[0016] This specification describes various embodiments of systems, apparatuses, devices, and methods for wireless communication related to model inference time and / or state determination.

[0017] Figure 1 A schematic diagram of an example wireless communication system 100 (also referred to herein as a mobile communication system) is shown, which includes multiple communication nodes (or nodes only) configured to communicate wirelessly with each other. Typically, the communication nodes include at least one user equipment 102 and at least one network device or base station 104. Figure 1The example wireless communication system 100 is shown as including two user equipments 102, including a first user equipment 102 (1) and a second user equipment 102 (2), and a network device 104. However, various other examples of wireless communication systems 100 including any of various combinations of one or more user equipments 102 and / or one or more network devices 104 are possible.

[0018] Generally, user equipment described herein, such as user equipment 102, may include a single electronic device or apparatus capable of wireless communication over a network, or multiple electronic devices or apparatuses (e.g., a network thereof). User equipment may include, or be otherwise referred to as, a user terminal, user terminal equipment, or user equipment (UE). Furthermore, user equipment may be, or includes, mobile devices (such as mobile phones, smartphones, smartwatches, tablets, laptops, vehicles, or other vessels (human-, motor-, or engine-driven, such as cars, airplanes, trains, ships, or bicycles, as non-limiting examples)) or fixed or stationary devices (such as desktop computers or other computing devices that are not typically moved for extended periods, such as appliances, other relatively heavy devices including the Internet of Things (IoT), or computing devices used in commercial or industrial environments, as non-limiting examples). In various embodiments, user equipment 102 may include a transceiver circuitry 106 coupled to antenna 108 to enable wireless communication with network device 104. Transceiver circuitry 106 may also be coupled to processor 110, which may also be coupled to memory 112 or other storage devices. The memory 112 may store instructions or code that, when read and executed by the processor 110, enable the processor 110 to implement the various methods described herein.

[0019] Furthermore, in general, network devices as described herein, such as network device 104, may include a single electronic device or apparatus, or multiple electronic devices or apparatuses (e.g., a network thereof), and may include one or more wireless access nodes, base stations, or other wireless network access points capable of wirelessly communicating with one or more user equipments and / or one or more other network devices 104 via a network. For example, in various embodiments, network device 104 may include a 4G LTE base station, a 5G NR base station, a 5G central unit base station, a 5G distributed unit base station, a next-generation node B (gNB), an enhanced node B (eNB), or other similar or next-generation (e.g., 6G) base station. Network device 104 may include a transceiver circuitry 114 coupled to an antenna 116, which may include an antenna tower 118 in various methods to enable wireless communication with user equipment 102 or another network device 104. The transceiver circuitry 114 may also be coupled to one or more processors 120, which may also be coupled to a memory 122 or other storage device. The memory 122 may store instructions or code that, when read and executed by the processor 120, enable the processor 120 to implement one or more of the methods described herein.

[0020] In various embodiments, two communication nodes in wireless system 100 (such as user equipment 102 and network equipment 104, two user equipment 102 without network equipment 104, or two network equipment 104 without user equipment 102) can be configured to wirelessly communicate with each other in or through a mobile network and / or radio access network according to one or more standards and / or specifications. Generally, these standards and / or specifications can define rules or procedures for communication nodes to wirelessly communicate, and in various embodiments, these rules or procedures may include rules or procedures for communication in the millimeter (mm) band and / or the use of multi-antenna schemes and beamforming capabilities. Alternatively or additionally, these standards and / or specifications are standards and / or specifications defining radio access technologies and / or cellular technologies, such as fourth-generation (4G) Long Term Evolution (LTE), fifth-generation (5G) New Radio (NR), or New Radio Unlicensed (NR-U), as non-limiting examples.

[0021] Furthermore, in the wireless system 100, communication nodes are configured to wirelessly transmit signals to each other. Typically, communication between two communication nodes in the wireless system 100 can be or includes transmission or reception, and is usually simultaneous, depending on the perspective of the specific node in the communication. For example, for a given communication between a first node and a second node, where the first node is transmitting a signal to the second node and the second node is receiving a signal from the first node, the first node can be referred to as a source or transmitting node or device, and the second node can be referred to as a destination or receiving node or device; this communication can be viewed as a transmission by the first node and a reception by the second node. Of course, since communication nodes in the wireless system 100 can both transmit and receive signals, a single communication node can simultaneously be both a transmitting / source node and a receiving / destination node, or switch between being a source / transmitting node and a destination / receiving node.

[0022] Furthermore, specific signals can be characterized or defined as uplink (UL) signals, downlink (DL) signals, or sidelink (SL) signals. An uplink signal is a signal transmitted from user equipment 102 to network device 104. A downlink signal is a signal transmitted from network device 104 to user equipment 102. A sidelink signal is a signal transmitted from one user equipment 102 to another user equipment 102, or from one network device 104 to another network device 104. Moreover, for sidelink transmission, the first / source user equipment 102 transmits the sidelink signal directly to the second / destination user equipment 102 without forwarding it to network device 104. Similarly, the first / source network device 104 transmits the sidelink signal directly to the second / destination network device 104 without forwarding it to user equipment 102.

[0023] Furthermore, at least some signals transmitted between communication nodes in system 100 can be characterized or defined as data signals or control signals. Generally, data signals are signals that include or carry data, such as multimedia data (e.g., voice and / or image data), while control signals are signals that carry control information that configures communication nodes in a specific way to communicate with each other, or otherwise controls how communication nodes transmit data signals to each other. Additionally, certain signals can be defined or characterized by combinations of data / control and uplink / downlink / sidelink, including uplink control signals, uplink data signals, downlink control signals, downlink data signals, sidelink control signals, and sidelink data signals.

[0024] For at least some specifications, such as 5G NR, data and control signals are transmitted and / or carried on physical channels. Typically, a physical channel corresponds to a set of time-frequency resources used for signal transmission. Different types of physical channels can be used to transmit different types of signals. For example, physical data channels (or data-only channels), also referred to herein as flow channels, are used to transmit data signals, and physical control channels (or control-only channels) are used to transmit control signals. Example types of flow channels (or physical data channels) include, but are not limited to, the Physical Downlink Shared Channel (PDSCH) for transmitting downlink data signals, the Physical Uplink Shared Channel (PUSCH) for transmitting uplink data signals, and the Physical Sidelink Shared Channel (PSSCH) for transmitting sidelink data signals. Furthermore, example types of physical control channels include, but are not limited to, the Physical Downlink Control Channel (PDCCH) for transmitting downlink control signals, the Physical Uplink Control Channel (PUCCH) for transmitting uplink control signals, and the Physical Sidelink Control Channel (PSCCH) for transmitting sidelink control signals. For simplicity, as used herein, unless otherwise stated, a specific type of physical channel is also used to refer to signals transmitted on that specific type of physical channel and / or transmissions on that specific type of transmission. As an example, PDSCH refers to the Physical Downlink Shared Channel itself, downlink data signals transmitted on the PDSCH, or downlink data transmissions. Therefore, a communication node transmitting or receiving a PDSCH implies that the communication node is transmitting or receiving signals on the PDSCH.

[0025] Furthermore, for at least some specifications, such as 5G NR, and / or for at least some types of control signals, the control signals transmitted by the communication nodes may include control information that includes information required to enable the transmission of one or more data signals between the communication nodes and / or to schedule one or more data channels (or one or more transmissions on a data channel). For example, such control information may include information required for the correct reception, decoding, and demodulation of data signals received on a physical data channel during data transmission, and / or information required for uplink scheduling authorization that informs the user equipment of resources and transmission formats for uplink data transmission. In some embodiments, the control information includes downlink control information (DCI) transmitted from network device 104 to user equipment 102 in the downlink direction. In other embodiments, the control information includes uplink control information (UCI) transmitted from user equipment 102 to network device 104 in the uplink direction, or sidelink control information (SCI) transmitted from one user equipment 102 (1) to another user equipment 102 (2) or from one network device 104 (1) to another network device 104 (2) in the sidelink direction.

[0026] Figure 2 This is a flowchart of an example method 200 for wireless communication related to model inference time. At block 202, user equipment 102 transmits the nominal inference times of M models to base station 104, where M is an integer of 2 or greater. At block 204, user equipment 102 receives an indication of a subset of models from the M models activated by base station 104. This subset includes N models activated by base station 104, where N is an integer of 1 or greater, and N is less than M. At block 206, user equipment 102 determines the actual inference time of the subset based on the activated model subset and the nominal inference time.

[0027] Figure 3 This is a flowchart of another example method 300 for wireless communication related to model inference time. At block 302, base station 104 receives the nominal inference times of M models from user equipment, where M is an integer of 2 or greater. At block 304, base station 104 activates a subset of models from the M models. This subset includes N models activated by the base station, where N is an integer of 1 or greater, and N is less than M. At block 306, base station 104 transmits an indication of the activated model subset to user equipment 102.

[0028] Figure 4 This is a flowchart of an example method 400 for wireless communication related to model states. At block 402, user equipment 102 determines one of a plurality of states of the model. At block 404, user equipment 102 determines, based on the state, whether to perform at least one operation associated with the model, wherein the at least one operation includes at least one of the following: model inference or model monitoring.

[0029] In some implementations of methods 200, 300, and / or 400, the actual inference time is determined based on N models of the nominal inference time and a subset of base station activations.

[0030] In some implementations of methods 200, 300, and / or 400, the actual inference time is determined based on the nominal inference time, N models of a subset of base station activations, and additional time units.

[0031] In some implementations of methods 200, 300, and / or 400, the actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determined according to the following formula:

[0032] ,

[0033] Where T i is the nominal inference time of the i-th model, and N is the number of models in the subset. It is an additional time unit, where i is an integer, where .

[0034] In some implementations of methods 200, 300, and / or 400, actual inference is determined based on nominal inference time, the number of models in the subset, the total artificial intelligence processing units (APUs) of user device 102, and the APU requirements of each model in the subset.

[0035] In some implementations of methods 200, 300, and / or 400, the actual inference time includes the i-th actual inference time of the i-th model. And the i-th actual inference time Determined according to the following formula:

[0036] ,or

[0037] ,

[0038] Among them, T i This is the nominal inference time of the i-th model. It is the total APU of the user equipment. This is the APU requirement for the i-th model. It is the total APU requirement of a subset of models, where i is an integer, where .

[0039] In some implementations of methods 200, 300, and / or 400, the total APU requirement of user equipment 102 and the APU requirement of the i-th model are reported by user equipment 102, indicated by base station 104, or predefined.

[0040] In some implementations of methods 200, 300, and / or 400, the actual inference time is determined based on the nominal inference time, the N models activated by base station 104, and the nominal number of activated models.

[0041] In some implementations of methods 200, 300 and / or 400, the actual inference time is also determined based on additional time units.

[0042] In some implementations of methods 200, 300, and / or 400, the actual inference time includes the i-th actual inference time of the i-th model. And the i-th actual inference time Determined according to the following formula:

[0043] ,

[0044] Where T i is the nominal inference time of the i-th model, and N is the number of models in the subset. It is an additional time unit, where i is an integer, where P is the nominal number of activated models.

[0045] In some implementations of methods 200, 300, and / or 400, the actual inference time is determined based on the nominal inference time and the number of layers.

[0046] In some implementations of methods 200, 300, and / or 400, the actual inference time includes the i-th actual inference time of the i-th model. And the i-th actual inference time Determined according to the following formula:

[0047] ,

[0048] Where T i L is the nominal inference time of the i-th model, where L is the number of layers and i is an integer. .

[0049] In some implementations of methods 200, 300, and / or 400, the actual inference time is determined based on the maximum number of layers.

[0050] In some implementations of methods 200, 300, and / or 400, the actual inference time includes the i-th actual inference time of the i-th model. And the i-th actual inference time Determined according to the following formula:

[0051] ,

[0052] Where T i L is the nominal inference time of the i-th model, where L is the number of layers. max It is the maximum number of layers, where i is an integer. .

[0053] In some implementations of methods 200, 300, and / or 400, the actual inference time is determined based on the nominal inference time and the number of time instances of the model input.

[0054] In some implementations of methods 200, 300, and / or 400, the actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determined according to the following formula:

[0055] ,

[0056] Where T i Let I be the nominal inference time of the i-th model, where I is the number of time instances and i is an integer. .

[0057] In some implementations of methods 200, 300, and / or 400, the actual inference time is determined based on the maximum number of time instances of the model input.

[0058] In some implementations of methods 200, 300, and / or 400, the actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determined according to the following formula:

[0059] ,

[0060] Where T i Let I be the nominal inference time of the i-th model, and let I be the number of time instances. max It is the maximum number of time instances, where i is an integer. .

[0061] In some embodiments of methods 200, 300, and / or 400, user equipment 102 receives and / or base station 104 transmits a command. This command instructs the activation of a subset of models. Additionally, user equipment 102 sends and / or base station 104 receives a report of the actual inference time for the subset. In any of the various embodiments, this command may be the same as, different from, or separate from the instruction received by user equipment 102 at block 204 and / or the instruction transmitted by base station 104 at block 304.

[0062] In some embodiments of methods 200, 300, and / or 400, the actual inference time is within the effective duration. Effective within a certain period. In some of these implementations, the effective duration is... As instructed by user equipment 102 or configured by base station 104.

[0063] In some implementations of methods 200, 300, and / or 400, the command indicates the expected inference time for a subset of the models.

[0064] In some embodiments of methods 200, 300, and / or 400, user equipment 102 transmits and / or base station 104 receives an evaluation report indicating whether the expected inference time can be met.

[0065] In some embodiments of methods 200, 300, and / or 400, user equipment 102 sends a report of a subset of the actual inference time according to a reporting period indicated by base station 104. Alternatively, base station 104 indicates a reporting period to user equipment 102, and user equipment 102 reports the actual inference time according to that reporting period.

[0066] In some embodiments of methods 200, 300, and / or 400, the reporting period is equal to the effective duration T of the actual inference time. valid .

[0067] In some embodiments of methods 200, 300 and / or 400, user equipment 102 transmits and / or base station 104 receives the effective duration and actual inference time together according to the reporting period.

[0068] In some implementations of methods 200, 300, and / or 400, multiple states of the model include: an active state, wherein the user device 102 is in an active state during the inference cycle. In-process model inference, and during the activation state monitoring cycle. Internal execution model monitoring; deactivation state, wherein user device 102 does not perform model inference and does not perform model monitoring, or wherein the user device is in a state longer than the activation state monitoring period. Deactivation status monitoring cycle The system performs model monitoring within the active state; and in an inactive state, where user equipment 102 performs model monitoring instead of model inference, wherein model monitoring is performed for a period longer or shorter than the active state monitoring period. Inactive status monitoring cycle Execution occurs within the active state; or model inference is performed along with model monitoring, where model inference occurs within a period longer than the active state inference cycle. Inactive state inference cycle It is executed internally.

[0069] In some embodiments of methods 200, 300 and / or 400, user equipment 102 receives a model pattern, wherein the model pattern includes a pattern period and an activity duration, and wherein the activity duration is less than the pattern period.

[0070] In some embodiments of methods 200, 300 and / or 400, user equipment 102 receives model patterns from base station 104, and / or base station 104 transmits model patterns to user equipment 102.

[0071] In some implementations of methods 200, 300, and / or 400, the activity duration begins at the start of each mode cycle or at an offset relative to the start of each mode cycle.

[0072] In some implementations of methods 200, 300, and / or 400, the model is active during the active duration and inactive during the inactive duration.

[0073] In some embodiments of methods 200, 300, and / or 400, a timer control model is used. In some of these embodiments, user equipment 102 restarts the timer in response to a timer restart condition being met. Alternatively, in some of these embodiments, the timer is configured by the base station.

[0074] In some implementations of methods 200, 300, and / or 400, the timer restart condition includes at least one of the following: the user equipment receives a command to activate the model; the user equipment receives a command to perform model monitoring; the user equipment reports model monitoring results; the user equipment reports inference time; the user equipment reports a model activation report; or the user equipment reports inference results.

[0075] In some embodiments of methods 200, 300, and / or 400, in response to the timer restart condition not being met within a single time unit, the user equipment decrements the timer at the end of the single time unit. In some of these embodiments, a single time unit includes: a frame, a subframe, a time slot, a sub-time slot, a second, a millisecond, or a period.

[0076] In some implementations of methods 200, 300, and / or 400, user equipment 102 switches the model to an inactive state when the timer expires.

[0077] In some implementations, user equipment 102 indicates a first group of one or more models and a second group of one or more models to base station 104. In any of these implementations, user equipment 102 indicates the first and second groups in conjunction with or independently of actions performed in relation to multiple states, such as actions performed at block 402 and / or block 404.

[0078] In some implementations of methods 200, 300 and / or 400, the first set is available at user equipment 102 and the second set is available at the server of user equipment 102.

[0079] In some implementations of methods 200, 300, and / or 400, the first group requires a first time delay so that the user device 102 can activate one or more models of the first group, and the second group requires a first time delay plus a second time delay so that the user device 102 can activate one or more models of the second group, wherein the first time delay is used for the user device 102 to prepare for model activation, and the second time delay is used for the user device 102 to download one or more models of the second group.

[0080] In some embodiments of methods 200, 300, and / or 400, user equipment 102 receives indications from base station 104 of a first group of one or more models and a second group of one or more models. In some of these embodiments, the first group is configured by base station 104, the second group is indicated by downlink control information (DCI) or media access control (MAC) control element (CE), the second group is selected from the first group of models, and / or base station 104 indicates to the user equipment that at least one of the one or more models in the second group is activated.

[0081] Other methods are also possible, including, but not limited to, methods that combine one or more blocks or actions from each of two or more of methods 200, 300, and 400, and / or methods that include fewer than all actions performed for the above-described embodiments of methods 200, 300, and 400. For example, some other methods may convey instructions for one or more models of the first and second groups independently of or not necessarily by determining one or more states of the model. Various methods or combinations of methods based on multiple embodiments of methods 200, 300, and 400 are possible.

[0082] Further details of the actions performed by the communication nodes in the wireless communication system 100 will now be described, wherein any or all of the actions may be implemented in any of the various implementations of method 200, method 300, method 400 and / or other methods.

[0083] Some implementations of the wireless communication system 100 may utilize artificial intelligence (AI) and / or machine learning (ML) (collectively referred to herein as AI / ML) to enhance performance and / or operation. For example, implementing AI / ML techniques in a wireless communication system can improve system operating efficiency, such as by reducing the overhead of reference signals through AI / ML inference and prediction.

[0084] Furthermore, in some embodiments of the AI / ML technology implemented in the wireless communication system 100, the wireless communication system 100 includes or employs an AI / ML model that performs one or more functions or operations (such as inference). Typically, a model can refer to a functionality, function, functional module, processing method, information processing method, implementation method, feature, feature group, configuration, configuration set, dataset (e.g., for model training), or data-driven algorithm. Typically, these models are executed, computed, or processed by the user equipment 102 (e.g., a user equipment (UE)), but in any of the various embodiments, they may also be executed at the base station 104.

[0085] Furthermore, in any of the various implementations, a model (as used herein, unless otherwise explicitly stated, the term model is used to refer to and / or interchangeably with AI model, ML model, or AI / ML model) can be a data-driven algorithm that applies AI / ML techniques to generate a set of outputs based on a set of inputs. Alternatively, a model can be a linear or nonlinear algorithm, or a combination of both. Additionally, functionality may refer to a feature enabled by an AI / ML model. Alternatively, functionality may refer to a set of parameters or configurations of a feature. For example, user equipment 102 may include, employ, utilize, or implement a convolutional neural network (CNN) model to predict the beam of communication; the CNN model is the model, and beam prediction is the functionality. Different models and / or functionalities may be associated with different configurations (e.g., Radio Resource Control (RRC) configurations). Furthermore, as used herein, model activation may refer to activating a corresponding configuration of user equipment 102. Similarly, model deactivation, switching, and fallback may refer to deactivating a corresponding configuration, switching a configuration, and falling back to a configuration without a model, respectively.

[0086] Furthermore, in some implementations, the AI ​​model can perform one or more functions, such as model training and model inference. Performing these functions consumes computational power. The methods described herein include coordinating computational power among communication nodes in a wireless communication system 100, including in cases where computational power is insufficient in one or more communication nodes.

[0087] In some implementations or scenarios, different user equipment 102 may be equipped with different computing capabilities. Alternatively, different models with varying complexities within the same user equipment 102 may require different computing capabilities. However, reporting the computing capabilities of each user equipment 102 may undesirably include disclosure of the user equipment 102's implementation design details. To avoid such reporting, user equipment 102 may report the inference time of each model to the base station.

[0088] In some implementations, user equipment 102 may report the nominal inference times of M models to a network device or base station 104. In response, base station 104 may activate a subset of the M models to user equipment 102. As used herein, unless explicitly stated otherwise, this subset is described as comprising N models, where M and N are integers, each of M and N is 1 or greater, and N is not greater than M. In some of these implementations, M is 2 or greater, and M is greater than N. User equipment 102 may determine the actual inference time based on the activated models and nominal inference times of the subset.

[0089] Furthermore, in some embodiments, the nominal inference time refers to the inference time when all the computing power of a given user device 102 is allocated to a model. Therefore, the more models activated, the longer the inference time required by the user device. In any of the various embodiments, the actual inference time is determined according to the following scheme.

[0090] In the first scheme, the actual inference time is determined based on the nominal inference time and the number N of activated models. In some implementations of the first scheme, it is assumed that the nominal inference time of the i-th model is represented as T. i The actual inference time of the i-th model is expressed as N models are activated. In some implementations, the actual inference time of the i-th model is... Determined according to the following mathematical formula: , where i is an integer, such that For example, if user device 102 reports an inference time of 1 millisecond (ms) for a model, this report could indicate that if all of the user device's computing power were allocated to the model, then user device 102 would need 1 ms to process model inference. Continuing this example, if two models are activated (i.e., N=2), then computing power is allocated to both models, and in turn, user device 102 would need twice the processing time (i.e., 2 ms in this case) to process model inference.

[0091] In the second scheme, the actual inference time is determined based on the nominal inference time, the number of activated models N, and additional time units. In some implementations of the second scheme, when N models are activated, the actual inference time of the i-th model is... Determined according to the following mathematical formula: T i is the nominal inference time of the i-th model, and N is the number of activated models. It is an additional time unit, where i satisfies The integer. Furthermore, in some embodiments of the second scheme, the additional time unit may be configured by base station 104, may be reported by user equipment 102, or may be predefined.

[0092] In the third scheme, the actual inference time is determined based on the nominal inference time, the number of activated models N, and additional time units. In some implementations of the third scheme, where N models are activated, the actual inference time of the i-th model is... Determined according to the following mathematical formula: T i is the nominal inference time of the i-th model, and N is the number of activated models. It is an additional time unit, i is the one that satisfies The integer. For illustration, assume that user equipment 102 reports 1 ms of inference time for one model and an additional 0.5 ms of time. This reporting indicates, or signifies, that if all of the UE's computing power is allocated to one model, user equipment 102 requires 1 ms to process model inference. Furthermore, if two models are active (i.e., N=2), computing power is allocated to both models. Accordingly, user equipment 102 requires a longer inference time of 1.5 ms (i.e., ...). The reasoning is handled by the base station 104. Furthermore, in some implementations of the third scheme, the additional time unit can be configured by the base station 104, reported by the user equipment 102, or predefined.

[0093] In the fourth scheme, the actual inference time is determined based on the nominal inference time, the number of activated models N, the total AI processing units (APUs), and the APU requirements for each model. In some implementations of the fourth scheme, the actual inference time of the i-th model... The following mathematical formula is used to determine: if the nominal inference time of the i-th model is T... i And if N models are activated, then the actual inference time of this model is... T i is the nominal inference time of the i-th model, and N is the number of activated models. It is the total APU of user equipment 102. This is the APU requirement for the i-th activated model. It is the total APU requirement of N activation models, where i is the number of models that meet the requirements. Integers. In some of these implementations, the APU requirement of the i-th model. The hardware and / or software resource (e.g., computing power) requirements used to represent the i-th model. Additionally, the total APU... This represents the hardware and / or software resource (e.g., computing power) capabilities of user equipment 102. Furthermore, in any of the various implementations, the total APU of the i-th model... and / or APU requirements It can be reported by user equipment 102, indicated by base station 104, or predefined.

[0094] Furthermore, in some implementations, the nominal inference time refers to the inference time when the computing power of user equipment 102 is sufficient to power all a specific number of active models. However, if more than a specific number of additional models need to be activated, the computing power requirement for the total number of active models (including the specific number and one or more additional models) may exceed the capacity of user equipment 102. In turn, user equipment 102 may require a longer inference time.

[0095] The following describes additional schemes that can be used to determine actual inference time, including in cases where multiple activation models require computational capabilities beyond those of user device 102.

[0096] In the fourth scheme, the actual inference time is determined based on the nominal inference time, the number of activated models N, and the nominal number of activated models. In some implementations of the fourth scheme, the actual inference time of the i-th model... Determined according to the following mathematical formula:

[0097] ,

[0098] Where T i Let N be the nominal inference time of the i-th model, N be the number of activated models, P be the nominal number of activated models, and i be the number of models that satisfy the condition. The inference time is an integer. In practice, if the number of activated models N exceeds the nominal number of models P, more inference time is required. For illustration, suppose user device 102 reports an inference time of 1 ms for one model and reports a nominal number of activated models of 2. Accordingly, if only one model is activated, the inference time for one activated model is 1 ms. However, if, for example, three models are activated, the inference time is 1.5 ms. .

[0099] In the fifth scheme, the actual inference time is determined based on the nominal inference time, the number of activated models N, the nominal number of activated models, and additional time units. In some implementations of the fifth scheme, the actual inference time of the i-th model... Determined using the following mathematical formula:

[0100] ,

[0101] Where T i is the nominal inference time of the i-th model, N is the number of activated models, and P is the nominal number of activated models. It is an additional time unit, i is the one that satisfies The inference time is an integer. In practice, if the number of activated models exceeds the nominal number of models, more inference time is required. For illustration, suppose user device 102 reports an inference time of 1 ms for one model and reports a nominal number of activated models of 2. Accordingly, if only one model is activated, the inference time for that single model is 1 ms. However, if, for example, three models are activated, an inference time of 1.5 ms is used, assuming... It takes 0.5ms.

[0102] In the sixth scheme, the actual inference time is determined based on the nominal inference time, the number of activated models N, the total APU, and the APU requirement for each model. In some implementations of the sixth scheme, the actual inference time of the i-th model... Determined according to the following mathematical formula:

[0103] .

[0104] Where T i is the nominal inference time of the i-th model, and N is the number of activated models. It is the total APU of user equipment 102. This is the APU requirement for the i-th activated model. It is the total APU requirement of N activation models, where i is the number of models that meet the requirements. Integers.

[0105] Furthermore, in some implementations, user equipment 102 can determine the actual inference time based on the input dimension. For example, if the number of inputs is relatively large, user equipment 102 may require a longer inference time.

[0106] The following describes other schemes that can be used to determine actual inference time.

[0107] In the seventh scheme, the actual inference time is determined based on the nominal inference time and the number of layers. In some implementations of the seventh scheme, the actual inference time of the i-th model... Determined according to the following mathematical formula: T i Let L be the nominal inference time of the i-th model, L be the number of layers, and i be the number of layers that satisfy the condition. The nominal inference time is an integer. In some of these implementations, the nominal inference time is determined based on one layer. For example, the layer may be or refer to a multiple-input multiple-output (MIMO) layer or a model layer. Alternatively, in some implementations, the number of layers L is configured by the base station 104.

[0108] In the eighth scheme, the actual inference time is determined based on the nominal inference time, the number of layers, and the maximum number of layers. In some implementations of the eighth scheme, the actual inference time of the i-th model... Determined according to the following mathematical formula: T i L is the nominal inference time of the i-th model, where L is the number of layers. max It is the maximum number of layers, and i is the number of layers that satisfies... The integer. In the eighth scheme, the nominal inference time is effectively determined based on the maximum number of layers. In some implementations, the layers in the eighth scheme may refer to MIMO layers or model layers.

[0109] In the ninth scheme, the actual inference time is determined based on the nominal inference time and the number of time instances of the model input. In some implementations of the ninth scheme, the actual inference time of the i-th model... Determined according to the following mathematical formula: T i Let I be the nominal inference time of the i-th model, I be the number of time instances of the model input, and i be the number of time instances that satisfy the condition. The integer I. Under the ninth scheme, the nominal inference time can effectively be determined based on a time instance. To illustrate the use of Channel State Information (CSI) prediction as an example, if user equipment 102 uses or applies five historical CSIs as model inputs in the time domain, then I equals 5. Furthermore, in some implementations of the ninth scheme, the number of time instances I can be indicated or configured by base station 104.

[0110] In the tenth scheme, the actual inference time is determined based on the nominal inference time, the number of time instances of the model input, and the maximum number of time instances of the model input. In some implementations of the tenth scheme, the actual inference time of the i-th model... Determined according to the following mathematical formula: T i I is the nominal inference time of the i-th model, and I is the number of time instances of the model input. max It is the maximum number of time instances, where i is the number of instances that satisfy... Integers. Under the tenth scheme, the nominal inference time can be effectively determined based on the maximum number of time instances of the model input.

[0111] Furthermore, in some implementations, base station 104 can determine whether the model is activated or deactivated. However, the inference time of the model on the user equipment (e.g., UE) side can be determined by user equipment 102. In this case, some information exchange can be performed between user equipment 102 and base station 104 to ensure their understanding is consistent.

[0112] Furthermore, in some embodiments, base station 104 may send a command to user equipment 102, wherein the command instructs user equipment 102 to activate one or more models. In response to receiving the command, user equipment 102 may evaluate the inference time of one or more models to be activated and report the inference time to base station 104. In other words, user equipment 102 may report the inference time of one or more models to be activated. In some of these embodiments, the inference time reported by user equipment 102 may be at least within a certain duration. Effective within a certain period. Specifically, in these implementations, the duration... It can be indicated by user equipment 102 or configured by base station 104.

[0113] Furthermore, in some implementations, after receiving a report from user equipment 102, base station 104 may determine whether to activate one or more models based on inference time. If base station 104 determines that one or more models (or a subset of these models) are activated, user equipment 102 may begin activating the corresponding models. However, if base station 104 determines that one or more models (or a subset of these models) are not activated, user equipment 102 may stop activating the corresponding models.

[0114] If user equipment 102 has previously activated some models, in response to receiving a command to activate one or more models, user equipment 102 can report the inference time of the previously activated models and the inference time of the one or more models to be activated to base station 104. This is because if user equipment 102 activates one or more models indicated by the command, the inference time of the previously activated models may be affected.

[0115] Furthermore, in some embodiments, base station 104 may send a command to user equipment 102, wherein the command instructs user equipment 102 to activate one or more models, and the command indicates the expected inference time of one or more models. In some of these embodiments, in response to receiving the command, user equipment 102 may evaluate the inference time of one or more models to be activated and report to base station 104 whether the expected inference time can be guaranteed. Particularly in these embodiments where user equipment 102 sends an evaluation report to base station 104, the evaluation report includes at least one of the following: the expected inference time can be met; or the expected inference time cannot be met. Even if the expected inference time can be met, user equipment 102 may report to base station 104 that the model is being activated. Furthermore, if the expected inference time cannot be met, user equipment 102 may report to base station 104 that the model is not activated, or may report to base station 104 that the model is deactivated.

[0116] Furthermore, in some embodiments, base station 104 may configure or instruct user equipment 102 to report a period. In response, user equipment 102 may report inference time to base station 104 after or according to this period. Because inference time can be affected by various factors, such as overheating, the number of active models, memory resource shortages, etc., it may be beneficial for user equipment 102 to periodically report the inference time of active models. Alternatively, in some embodiments, the inference time reported by user equipment 102 may be at least for the effective duration. Effective within a certain period. In any of these implementations, the effective duration is... The period can be indicated by user equipment 102, or it can be the same as the period configured for base station 104. Alternatively, in some implementations, user equipment 102 can report the inference time and effective duration to the base station after this period. .

[0117] In some implementations, each model may be in different states. That is, at a given point in time, each model can be in one of multiple states and can be in different states at different points in time. It is possible to implement models in different states at different times to minimize computational power consumption. Specifically, in these implementations, a given model may be in one of three states.

[0118] The first state is the active state. In the active state, user equipment 102 can perform model inference and model monitoring. In some implementations, in the active state, the user equipment can perform inference over a specified period. In-process model inference can be performed, and the model monitoring cycle can be active. Internal execution model monitoring.

[0119] The second state is the deactivation state. In the deactivation state, user equipment 102 may not perform model inference. Furthermore, in some embodiments, user equipment 102 may not perform model monitoring in the deactivation state. In other embodiments, user equipment 102 may perform model monitoring during the deactivation state model monitoring cycle. Internal execution model monitoring. For at least some of these implementations, the deactivation state model monitoring cycle. Longer than or greater than the monitoring period of the activation state model In other embodiments of these implementations, the state monitoring cycle is deactivated. Shorter than or less than the monitoring cycle of the activation state model .

[0120] The third state is the inactive state. In some implementations, user equipment 102 may not perform model inference in the inactive state. In other implementations, user equipment 102 may perform model inference in the inactive state. In these other implementations, user equipment 102 may perform model inference cycles in the inactive state. Internal execution model inference. In some of these other implementations, the inactive state model inference cycle... Longer than or greater than the inference cycle of the active state model ,Right now, In other embodiments of these other implementations, the inactive state model inference cycle Shorter than or less than the inference cycle of the active state model ,Right now, Furthermore, user equipment can perform model monitoring during inactive states, such as during inactive state model monitoring cycles. In some implementations, the monitoring period for the inactive state model is longer than or greater than the monitoring period for the active state model. ,Right now, In other implementations, the inactive state model monitoring period... Less than or shorter than the activation model monitoring period ,Right now, .

[0121] Furthermore, in some embodiments, base station 104 may configure or indicate at least one model mode to user equipment 102. In some of these embodiments, the model mode includes a mode period and an activity duration, wherein the activity duration is shorter than the mode period. The activity duration may begin at the start of each mode period or at an offset relative to the start of each mode period. In some of these embodiments, this offset is also configured by base station 104. During the activity duration, the model is in an active state. Outside of the activity duration, the model is in an inactive state. The duration outside of the activity duration may be referred to as the inactive duration.

[0122] Figure 5 This is a sequence diagram, illustrating an example of a model pattern. Figure 5 In the example, base station 104 can configure the mode period, activity duration, and offset to 100 ms, 50 ms, and 10 ms, respectively. In this case, the activity duration in the first period can start with a 10 ms offset and end accordingly at 60 ms. Other durations are inactive durations, including one duration extending from 0 ms to 10 ms, and another duration extending from 60 ms to 100 ms. The same pattern will be repeated in subsequent mode periods.

[0123] Alternatively, in some embodiments, base station 104 may be configured with a timer to control the model. In some of these embodiments, the timer may be implemented using a timer restart condition. When the timer restart condition is met, user equipment 102 may restart the timer. Specifically, in these embodiments, the timer restart condition may include at least one of the following: user equipment 102 receives a command to activate the model; user equipment 102 receives a command to perform model monitoring; user equipment 102 reports model monitoring results; user equipment 102 reports inference time; user equipment 102 reports a model activation report; or user equipment 102 reports inference results.

[0124] In some implementations, if the time restart condition is not met within a time unit, the user equipment 102 may decrement the timer at the end of the time unit. The time unit may be a frame, a subframe, a time slot, a sub-time slot, a second, a millisecond, or a period, as a non-limiting example of any of the various implementations. Alternatively, the time unit may be configured by the base station 104 in any of the various implementations.

[0125] Furthermore, in some embodiments, user equipment 102 can switch the model to an inactive state when the timer expires. In other embodiments, user equipment 102 can switch the model to a deactivated state when the timer expires.

[0126] Furthermore, in some implementations, base station 104 may send commands to user equipment 102, instructing user equipment 102 to switch one or more models between two states and / or to put one or more models in a specific state. For example, the command may instruct user equipment 102 to switch a model to a deactivated state, an inactive state, or an active state. Figure 6 A schematic diagram of the overall state transition of the model is shown.

[0127] Alternatively, in some implementations, user equipment 102 may indicate two sets of models to base station 104, including a first set and a second set. Each set may include one or more models. The first set may be available at user equipment 102. The second set may be available at a server (e.g., an OTT server that user equipment 102 can access and / or communicate with). To use one or more models in the second set, user equipment 102 may access the server to download the one or more models. In contrast, for models in the first set, user equipment 102 already possesses the model (e.g., has already downloaded the model), and therefore, user equipment 102 does not need to access the server to obtain the model. Accordingly, to use models in the second set, user equipment 102 may require additional time to download the model before user equipment 102 can activate it.

[0128] Alternatively, in some other embodiments, user equipment 102 may indicate two sets of models to base station 104, including a first set and a second set. Each set may include one or more models. In some of these embodiments, for models in the first set, user equipment 102 may generate a first time delay to activate the model. For models in the second set, user equipment 102 may generate a first time delay plus a second time delay to activate the model. In particular, in these embodiments, the first time delay is used for user equipment 102 to prepare for model activation. The second time delay is used for user equipment 102 to download one or more models.

[0129] Alternatively, in some implementations, base station 104 may indicate two sets of models to user equipment 102, including a first set and a second set. Each set may include one or more models. The first set may be configured by base station 104. The second set may be indicated by downlink control information (DCI) or media access control (MAC) control element (CE). Alternatively, the second set of models may be selected from the first set of models. Base station 104 may activate one or more models in the second set for user equipment 102, and / or indicate to user equipment 102 which(s) of the second set of models to activate.

[0130] Alternatively, in some implementations, the first group of models is in a deactivated state, and the second group of models is in an inactive state. In some of these implementations, base station 104 can activate one or more models in the second group, such as by switching the one or more models in the second group from an inactive state to an active state.

[0131] In some implementations or scenarios, user equipment 102 may not have sufficient computing power to perform one or more operations associated with the model, such as model training. In some of these implementations, user equipment 102 may request computing power from base station 104 to perform one or more operations. However, in order to perform the one or more operations, user equipment 102 may first acquire the dataset for the one or more operations. Taking model training as an example, user equipment 102 may first acquire the dataset used for model training.

[0132] Furthermore, in some implementations, user equipment 102 may send a dataset request to base station 104. This dataset request may include one or more of the following information: request identifier (ID); model input and output requirements; dataset resolution requirements; or dataset size requirements.

[0133] The request ID can be an index. When base station 104 sends back dataset information, it can also send the same request ID to user equipment 102. In turn, user equipment 102 can understand that the dataset information was used for a previous request. In some implementations, the request ID may also include the vendor ID of user equipment 102.

[0134] Model input and output requirements may include dataset dimensionality information for the input and output data. Alternatively, model input and output requirements may also include data quantization requirements. For example, user equipment 102 might require a 12x14 input dimension, representing modulation symbols in a resource block (RB) (e.g., 12 resource elements (REs)) within a time slot (e.g., 14 OFDM symbols). Quantization requirements may include linear quantization or vector quantization.

[0135] Dataset resolution requirements may include the expected resolution, such as 1 sample per 1ms, or numbers accurate to two decimal places.

[0136] Data size requirements may include the total size of the dataset or the total number of samples in the dataset.

[0137] In some implementations, in response to receiving a dataset request, base station 104 may send a dataset request response back to user equipment 102. Furthermore, in response to the dataset request, base station 104 may identify a dataset that meets the requirements of the request. If base station 104 identifies a dataset that meets the requirements, the dataset request response may include at least one of the following information about the dataset: a request identifier (ID); dataset information; a dataset address; or a dataset ID. In some implementations, the dataset information may include at least one of the following: dataset size, dimensions, resolution, input information, or output information. Additionally, the dataset address may indicate the location where the dataset is saved or stored in memory. In any of the various implementations, user equipment 102 and / or base station 104 may access the dataset via the dataset address. Furthermore, the dataset ID is an identifier or index of the dataset.

[0138] In some implementations, if base station 104 does not identify any dataset that meets the requirements, the dataset request response includes a request ID and an error feedback. The error feedback may include a reason for the error, indicating that the base station cannot identify any dataset that meets the request requirements.

[0139] In addition, in response to a data set request response from base station 104, user equipment 102 may send a report to base station 104 to indicate that a data set request response has been received and / or to indicate that data set information has been received.

[0140] Furthermore, in some embodiments, user equipment 102 and base station 104 can perform information exchange for model training. In some of these embodiments, user equipment 102 can send a model training request to base station 104, wherein the request includes at least one of the following: request ID; dataset ID; one or more model requirements; or computing power requirements.

[0141] In some implementations, the request ID is an index. When base station 104 sends back model training information, base station 104 may also send the same request ID to user equipment 102. In turn, user equipment 102 can understand that the model training information is for a previous request. In some implementations, the request ID may also include the vendor ID of user equipment 102. The dataset ID may identify the dataset used for model training. As a non-limiting example, model requirements may include at least one of the following: model structure, number of layers, or applicable conditions. As a non-limiting example, applicable conditions may include applicable region information, applicable frequency / band / carrier information, and / or applicable time information. Computational capability requirements may include the expected computational capability of the model and / or the maximum computational capability of the model. In some implementations, computational capability can be expressed in floating-point operations per second (FLOPs).

[0142] Furthermore, in some embodiments, base station 104 may send a model training response back to user equipment 102. If base station 104 successfully trains a model that meets the requirements, the model training response includes at least one of the following: a request ID, a model ID, model information, or computing power information. The model ID can identify the model. The model information may include the model itself, or the address where the model is stored in memory, which in turn allows user equipment 102 to download the model from memory. The computing power information may include the computing power required for the model to perform model inference. On the other hand, if base station 104 fails to train a model that meets the requirements, the model training response may include a request ID and error feedback. The error feedback may include the reason for the error, i.e., that base station 104 cannot train a model that meets the requirements.

[0143] In response to a model training response from base station 104, user equipment 102 may send a report to base station 104 to indicate that a model training response has been received and / or to indicate that model information has been received.

[0144] The above description and accompanying drawings provide specific example embodiments and implementations. However, the described subject matter can be embodied in a variety of different forms, and therefore, the covered or claimed subject matter is intended to be construed as not being limited to any of the example embodiments described herein. A reasonably broad scope is intended for the claimed or covered subject matter. In addition, for example, the subject matter can be embodied as a method, apparatus, component, system, or non-transitory computer-readable medium for storing computer code. Thus, for example, embodiments can take the form of hardware, software, firmware, storage media, or any combination thereof. For example, the above method embodiments can be implemented by executing computer code stored in memory, by components, apparatus, or a system including memory and a processor.

[0145] Throughout the specification and claims, terms may have subtle meanings implied or suggested in the context, beyond their explicitly stated meanings. Similarly, the phrase "in one embodiment / implementation" as used herein does not necessarily refer to the same embodiment, and the phrase "in another embodiment / implementation" as used herein does not necessarily refer to a different embodiment. For example, the claimed subject matter includes combinations of all or some of the exemplary embodiments.

[0146] Generally, terms can be understood, at least in part, from their usage in context. For example, terms used herein, such as “and,” “or,” or “and / or,” may include a variety of meanings, which may depend at least in part on the context in which these terms are used. Typically, “or,” if used to relate a list, such as A, B, or C, means A, B, and C, which is inclusive, and A, B, or C, which is exclusive. Furthermore, the term “one or more,” as used herein, depends at least in part on the context and can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, and characteristics in a plural sense. Similarly, terms such as “a,” “one,” or “the” can be understood to convey either singular or plural usage, depending at least in part on the context. Moreover, the term “based on” may be understood not necessarily to convey an exclusive set of factors, but may allow for the presence of other factors that are not necessarily explicitly described, depending at least in part on the context.

[0147] References to features, advantages, or similar language throughout this specification do not imply that all features and advantages achievable with this solution should be or are included in any single implementation thereof. Rather, references to features and advantages are understood to mean that a particular feature, advantage, or characteristic described in connection with an embodiment is included in at least one embodiment of this solution. Therefore, throughout this specification, discussions of features and advantages, and similar language, may, but do not necessarily, refer to the same embodiments.

[0148] Furthermore, the features, advantages, and characteristics described herein can be combined in one or more embodiments in any suitable manner. Based on the description herein, those skilled in the art will recognize that this solution can be practiced without one or more specific features or advantages of a particular embodiment. In other instances, additional features and advantages that may not be present in all embodiments of this solution may be recognized in certain embodiments.

[0149] The subject matter of this disclosure may also include the following:

[0150] The first aspect includes a wireless communication method comprising: transmitting, by a user equipment, nominal inference times of M models to a base station, wherein M is an integer of 2 or greater; receiving, by the user equipment, an indication of a subset of models activated by the base station, wherein the subset comprises N models activated by the base station, wherein N is an integer of 1 or greater and N is less than M; and determining, by the user equipment, the actual inference time of the subset based on the activated model subset and the nominal inference times.

[0151] The second aspect includes a wireless communication method comprising: receiving, by a base station, nominal inference times of M models from a user equipment, wherein M is an integer of 2 or greater; activating a subset of models from the M models by the base station, wherein the subset comprises N models activated by the base station, wherein N is an integer of 1 or greater and N is less than M; and transmitting, by the base station, an indication of the activated subset of models to the user equipment.

[0152] The third aspect includes either the first or the second aspect, and also includes N models in which the actual inference time is determined based on a subset of the nominal inference time and base station activation.

[0153] The fourth aspect includes either the first or the second aspect, and also includes determining the actual inference time based on N models of a subset of base station activations and additional time units, according to the nominal inference time.

[0154] The fifth aspect includes the fourth aspect, and also includes the actual inference time of the i-th model, which is the i-th actual inference time. Wherein, the i-th actual inference time Determined according to the following formula:

[0155] ,

[0156] Where T i is the nominal inference time of the i-th model, and N is the number of models in the subset. It is an additional time unit, where i is an integer, where .

[0157] The sixth aspect includes either the first or the second aspect, and also includes determining the actual inference time based on the nominal inference time, the number of models in the subset, the total AI processing units (APU) of the user device, and the APU requirements of each model in the subset.

[0158] The seventh aspect includes the sixth aspect, and also includes the actual inference time of the i-th model, which includes the i-th actual inference time. And the i-th actual inference time Determined according to the following formula:

[0159] ,or

[0160] ,

[0161] Where T i This is the nominal inference time of the i-th model. It is the total APU of the user equipment. This is the APU requirement for the i-th model. It is the total APU requirement of a subset of models, where i is an integer, where .

[0162] The eighth aspect includes either the sixth or seventh aspect, and also includes the total APU requirement of the user equipment and the APU requirement of the i-th model being reported by the user equipment, indicated by the base station, or predefined.

[0163] The ninth aspect includes either the first or the second aspect, and also includes determining the actual inference time based on the nominal inference time, the N models activated by the base station, and the number of nominally activated models.

[0164] The tenth aspect includes the ninth aspect, and further includes determining the actual inference time based on additional time units.

[0165] The eleventh aspect includes the tenth aspect, and also includes the actual inference time of the i-th model, which includes the i-th actual inference time. Wherein, the i-th actual inference time Determined according to the following formula:

[0166] ,

[0167] Where T i is the nominal inference time of the i-th model, and N is the number of models in the subset. It is an additional time unit, where i is an integer, where P is the nominal number of activated models.

[0168] The twelfth aspect includes either the first or the second aspect, and also includes determining the actual inference time based on the nominal inference time and the number of layers.

[0169] The thirteenth aspect includes the twelfth aspect, and also includes the actual inference time of the i-th model, which includes the i-th actual inference time. The i-th actual inference time Determined according to the following formula:

[0170] ,

[0171] Where T i L is the nominal inference time of the i-th model, where L is the number of layers and i is an integer. .

[0172] The fourteenth aspect includes the twelfth aspect, and also includes further determining the actual inference time based on the maximum number of layers.

[0173] The fifteenth aspect includes the fourteenth aspect, and also includes the actual inference time of the i-th model, which includes the i-th actual inference time. Wherein, the i-th actual inference time Determined according to the following formula:

[0174] ,

[0175] Where T i L is the nominal inference time of the i-th model, where L is the number of layers. max It is the maximum number of layers, where i is an integer. .

[0176] The sixteenth aspect includes either the first or the second aspect, and also includes determining the actual inference time based on the nominal inference time and the number of time instances of the model input.

[0177] The seventeenth aspect includes the sixteenth aspect, and also includes the actual inference time of the i-th model, which includes the i-th actual inference time. Wherein, the i-th actual inference time Determined according to the following formula:

[0178] ,

[0179] Where T i Let I be the nominal inference time of the i-th model, where I is the number of time instances and i is an integer. .

[0180] The eighteenth aspect includes the sixteenth aspect, and further includes determining the actual inference time based on the maximum number of time instances of the model input.

[0181] The nineteenth aspect includes the eighteenth aspect, and also includes the actual inference time of the i-th model, which includes the i-th actual inference time. Wherein, the i-th actual inference time Determined according to the following formula:

[0182] ,

[0183] Where T i Let I be the nominal inference time of the i-th model, and let I be the number of time instances. max It is the maximum number of time instances, where i is an integer. .

[0184] The twentieth aspect includes any one of the first to nineteenth aspects, and further includes: receiving a command from the base station by the user equipment, wherein the command instructs to activate a subset of the models; and sending a report of the actual inference time of the subset by the user equipment.

[0185] The twenty-first aspect includes any one of the first to twentieth aspects, and further includes: transmitting a command from the base station to the user equipment, wherein the command instructs the activation of a subset of models; and receiving a report of the actual inference time of the subset from the base station.

[0186] The twenty-second aspect includes either the twentieth or the twenty-first aspect, and also includes the actual reasoning time within the effective duration. Valid within, where the effective duration is... Instructed by the user equipment or configured by the base station.

[0187] The twenty-third aspect includes any one of aspects twenty to twenty-two, and also includes a provision in which the command further indicates the expected inference time for a subset of the model.

[0188] The 24th aspect includes the 23rd aspect, and also includes: transmitting an evaluation report from the user equipment to the base station, the evaluation report indicating whether the expected inference time can be met.

[0189] The twenty-fifth aspect includes any one of the twenty-third or twenty-fourth aspects, and further includes: receiving an evaluation report from the user equipment by the base station, the evaluation report indicating whether the expected inference time can be met.

[0190] The twenty-sixth aspect includes any one of aspects twentieth to twenty-fifth, and also includes a report in which the user equipment transmits a subset of the actual inference time according to the reporting period indicated by the base station.

[0191] The twenty-seventh aspect includes any one of aspects twenty to twenty-six, and also includes the fact that the base station instructs the user equipment to report a reporting period, and the user equipment reports the actual inference time according to the reporting period.

[0192] The twenty-eighth aspect includes either the twenty-sixth or twenty-seventh aspect, and also includes the effective duration T in which the reporting period is equal to the effective inference time. valid .

[0193] The 29th aspect includes the 28th aspect, and also includes: the user equipment transmitting the effective duration and actual inference time together according to the reporting cycle.

[0194] The thirtieth aspect includes either the twenty-eighth or twenty-ninth aspect, and also includes: the effective duration and actual inference time being received by the base station together according to the reporting period.

[0195] The thirty-first aspect includes a wireless communication method comprising: determining, by a user equipment, one of a plurality of states of a model; and determining, by the user equipment, whether to perform at least one operation associated with the model based on the state, wherein the at least one operation includes at least one of: model inference or model monitoring.

[0196] The thirty-second aspect includes the thirty-first aspect, and also includes multiple states including: an active state, wherein the user equipment is in an active state during the inference cycle. In-process model inference and during the activation state monitoring cycle. In-process model monitoring; deactivation state, where the user device does not perform model inference and does not perform model monitoring, or where the user device is in a state longer than the activation state monitoring period. Deactivation status monitoring cycle In-process model monitoring; and in an inactive state, where the user device: does not perform model inference but performs model monitoring, wherein model monitoring is performed at intervals longer or shorter than the active state monitoring period. Inactive status monitoring cycle Executed within the context of the active state; or performing model inference and model monitoring, wherein model inference occurs within a period longer than the active state inference cycle. Inactive state inference cycle It is executed internally.

[0197] The thirty-third aspect includes either the thirty-first or thirty-second aspect, and further includes: receiving a model pattern by a user equipment, the model pattern including a pattern period and an activity duration, wherein the activity duration is less than the pattern period.

[0198] The thirty-fourth aspect includes the thirty-third aspect, and also includes the model mode in which the user equipment receives data from the base station.

[0199] The thirty-fifth aspect includes any one of the thirty-third or thirty-fourth aspects, and also includes the fact that the duration of the activity begins at the start of each pattern cycle or at an offset relative to the start of each pattern cycle.

[0200] The thirty-sixth aspect includes either the thirty-third or the thirty-fourth aspect, and also includes the fact that the model is active during the active duration and inactive during the inactive duration.

[0201] The thirty-seventh aspect includes any one of aspects thirty-one to thirty-six, and also includes a timer control model configured by the base station therein, wherein the user equipment restarts the timer in response to the satisfaction of the timer restart condition.

[0202] The thirty-eighth aspect includes the thirty-seventh aspect, and also includes the timer restart condition including at least one of the following: the user equipment receives a command to activate the model; the user equipment receives a command to perform model monitoring; the user equipment reports the model monitoring results; the user equipment reports the inference time; the user equipment reports the model activation report; or the user equipment reports the inference results.

[0203] The thirty-ninth aspect includes any one of the thirty-seventh or thirty-eighth aspects, and further includes: in response to the failure to meet the timer restart condition within a single time unit, the user equipment decrements the timer at the end of the single time unit, wherein the single time unit includes: a frame, a subframe, a time slot, a sub-time slot, a second, a millisecond, or a period.

[0204] The fortieth aspect includes any one of aspects thirty-seven to thirty-nine, and also includes the user equipment switching the model to an inactive state when the timer expires.

[0205] The forty-first aspect includes any one of aspects thirty-first to forty-fifth, and also includes the user equipment indicating to the base station one or more models of the first group and one or more models of the second group.

[0206] The forty-second aspect includes the forty-first aspect, and also includes the first set of which is available at the user equipment and the second set of which is available at the server of the user equipment.

[0207] The forty-third aspect includes any one of the forty-first or forty-second aspects, and also includes the first group requiring a first time delay so that the user equipment can activate one or more models of the first group, and the second group requiring a first time delay plus a second time delay so that the user equipment can activate one or more models of the second group, wherein the first time delay is used for the user equipment to prepare for model activation, and the second time delay is used for the user equipment to download one or more models of the second group.

[0208] The forty-fourth aspect includes any one of aspects thirty-one to forty, and further includes a user equipment receiving from a base station an indication of a first group of one or more models and a second group of one or more models, wherein at least one of the following is satisfied: the first group is configured by the base station; the second group is indicated by downlink control information (DCI) or media access control (MAC) control element (CE); the second group is selected from the first group of models; or the base station indicates to the user equipment that at least one of the second group of one or more models is activated.

[0209] The forty-fifth aspect includes a wireless communication device comprising at least one processor and a memory, wherein the at least one processor is configured to cause the device to perform any one of the first to forty-fourth aspects.

[0210] The forty-sixth aspect includes a computer program product comprising a computer-readable program medium including code stored thereon, which, when executed by a processor, causes the processor to implement any one of the first to forty-fourth aspects.

[0211] In addition to the features mentioned in each of the individual aspects listed above, some examples may be shown individually or in combination of optional features mentioned in the dependency aspect and / or disclosed in the description above and shown in the figures.

Claims

1. A wireless communication method, the method comprising: The nominal inference time for the user equipment to transmit M models to the base station, where M is an integer of 2 or greater; The user equipment receives an indication of a subset of models from M models activated by the base station, wherein the subset includes N models activated by the base station, where N is an integer of 1 or greater and N is less than M; The user equipment determines the actual inference time of the subset based on the activated model subset and the nominal inference time.

2. A wireless communication method, the method comprising: The nominal inference time for the base station to receive M models from the user equipment, where M is an integer of 2 or greater; The base station activates a subset of models from the M models, wherein the subset comprises N models activated by the base station, where N is an integer of 1 or greater, and N is less than M; and The base station transmits an indication of the activated subset of models to the user equipment.

3. The method according to claim 1, wherein, The actual inference time is determined based on the nominal inference time and N models of the subset activated by the base station.

4. The method according to claim 1, wherein, The actual inference time is determined based on the nominal inference time, N models of the subset activated by the base station, and additional time units.

5. The method according to claim 4, wherein, The actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determine using the following formula: ， Where T i is the nominal inference time of the i-th model, and N is the number of models in the subset. It is the additional time unit, where i is an integer, where .

6. The method according to claim 1, wherein, The actual inference time is determined based on the nominal inference time, the number of models in the subset, the total AI processing units (APU) of the user device, and the APU requirements of each model in the subset.

7. The method according to claim 6, wherein, The actual inference time includes the i-th actual inference time of the i-th model. And the i-th actual inference time Determined according to the following formula: ,or ， Where T i This is the nominal inference time of the i-th model. It is the total APU of the user equipment. This is the APU requirement for the i-th model. It is the total APU requirement of the subset of the models, where i is an integer, where .

8. The method according to claim 6, wherein, The total APU requirement of the user equipment and the APU requirement of the i-th model are reported by the user equipment, indicated by the base station, or predefined.

9. The method according to claim 1, wherein, The actual inference time is determined based on the nominal inference time, the N models activated by the base station, and the nominal number of activated models.

10. The method according to claim 9, wherein, The actual inference time is also determined based on additional time units.

11. The method according to claim 10, wherein, The actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determine using the following formula: ， Where T i is the nominal inference time of the i-th model, and N is the number of models in the subset. It is the additional time unit, where i is an integer, where P is the nominal number of activated models.

12. The method according to claim 1, wherein, The actual inference time is determined based on the nominal inference time and the number of layers.

13. The method according to claim 12, wherein, The actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determine using the following formula: ， Where T i L is the nominal inference time of the i-th model, L is the number of layers, and i is an integer. .

14. The method according to claim 12, wherein, The actual inference time is also determined based on the maximum number of layers.

15. The method according to claim 14, wherein, The actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determine using the following formula: ， Where T i L is the nominal inference time of the i-th model, and L is the number of layers. max It is the maximum number of layers, where i is an integer. .

16. The method according to claim 1, wherein, The actual inference time is determined based on the nominal inference time and the number of time instances of the model input.

17. The method according to claim 16, wherein, The actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determine using the following formula: ， Where T i Let I be the nominal inference time of the i-th model, where I is the number of time instances and i is an integer. .

18. The method according to claim 16, wherein, The actual inference time is also determined based on the maximum number of time instances of the model input.

19. The method according to claim 18, wherein, The actual inference time includes the i-th actual inference time of the i-th model. Wherein, the i-th actual inference time Determine using the following formula: ， Where T i Let I be the nominal inference time of the i-th model, and let I be the number of time instances. max It is the maximum number of time instances, where i is an integer. .

20. The method according to claim 1, further comprising: The user equipment receives a command from the base station, wherein the command indicates activation of the model subset; and The user equipment sends a report on the actual inference time of the subset.

21. The method according to claim 2, further comprising: The base station transmits a command to the user equipment, wherein the command instructs the activation of the model subset; and The base station receives a report of the actual inference time of the subset.

22. The method according to claim 20 or 21, wherein, The actual inference time is within the effective duration. Valid within, wherein the effective duration Instructed by the user equipment or configured by the base station.

23. The method according to claim 20 or 21, wherein, The command also indicates the expected inference time for the subset of models.

24. The method of claim 23, further comprising: The user equipment transmits an evaluation report to the base station, the evaluation report indicating whether the expected inference time can be met.

25. The method of claim 23, further comprising: The base station receives an evaluation report from the user equipment, the evaluation report indicating whether the expected inference time can be met.

26. The method of claim 20, wherein, The user equipment sends a report of the actual inference time of the subset according to the reporting period indicated by the base station.

27. The method according to claim 21, wherein, The base station indicates a reporting period to the user equipment, and the user equipment will report the actual inference time according to the reporting period.

28. The method according to any one of claims 26 or 27, wherein, The reporting period is equal to the effective duration T of the actual inference time. valid .

29. The method of claim 28, further comprising: The user equipment transmits the effective duration and the actual inference time together according to the reporting period.

30. The method of claim 28, further comprising: The base station receives the effective duration and the actual inference time together according to the reporting period.

31. A wireless communication method, the method comprising: The user equipment determines one of the multiple states of the model; as well as The user equipment determines whether to perform at least one operation associated with the model based on the state, wherein the at least one operation includes at least one of the following: model inference or model monitoring.

32. The method according to claim 31, wherein, The plurality of states include: Activation state, wherein the user equipment is in the activation state inference cycle In-process model inference and during the activation state monitoring cycle. Internal execution model monitoring; In a deactivated state, wherein the user equipment does not perform model inference and does not perform model monitoring, or wherein the user equipment is in a monitoring period longer than the activation state. Deactivation status monitoring cycle Internal execution model monitoring; and Inactive state, wherein the user equipment: Instead of performing model inference, model monitoring is performed, wherein the model monitoring is performed at intervals longer or shorter than the activation state monitoring period. Inactive status monitoring cycle Executed internally; or Perform model inference and perform model monitoring, wherein the model inference is performed in a period longer than the inference cycle of the active state. Inactive state inference cycle It is executed internally.

33. The method of claim 31, further comprising: The user equipment receives a model pattern, which includes a pattern period and an activity duration, wherein the activity duration is less than the pattern period.

34. The method according to claim 33, wherein, The user equipment receives the model pattern from the base station.

35. The method according to any one of claims 33 or 34, wherein, The duration of the activity begins at the start of each pattern cycle or at an offset relative to the start of each pattern cycle.

36. The method according to any one of claims 33 or 34, wherein, The model is active during the activity duration and inactive during the inactive duration.

37. The method according to claim 31, wherein, The model is controlled by a timer configured by the base station, wherein the user equipment restarts the timer in response to the timer restart condition being met.

38. The method according to claim 37, wherein, The timer restart condition includes at least one of the following: The user equipment receives a command to activate the model; The user equipment receives a command to perform model monitoring; The user equipment reports the monitoring results of the model. The user equipment reports inference time; The user equipment report model activation report; or The user equipment reports the reasoning results.

39. The method of claim 37, wherein, in response to the timer restart condition not being met within a single time unit, the user equipment decrements the timer at the end of the single time unit, wherein, The single time unit includes: a frame, a subframe, a time slot, a sub-time slot, a second, a millisecond, or a period.

40. The method of claim 37, wherein, When the timer expires, the user equipment switches the model to an inactive state.

41. The method according to claim 31, wherein, The user equipment indicates one or more models in the first group and one or more models in the second group to the base station.

42. The method according to claim 41, wherein, The first set is available at the user equipment, and the second set is available at the server of the user equipment.

43. The method according to claim 41, wherein, The first group requires a first time delay so that the user equipment can activate one or more models in the first group. The second group requires the first time delay plus a second time delay so that the user equipment can activate one or more models in the second group. The first time delay is used for the user equipment to prepare for model activation, and the second time delay is used for the user equipment to download one or more models in the second group.

44. The method according to claim 31, wherein, The user equipment receives indications from the base station for a first group of one or more models and a second group of one or more models, wherein at least one of the following is satisfied: The first group is configured by the base station; The second group is indicated by downlink control information (DCI) or media access control (MAC) control element (CE); The second group was selected from the first group of models; or The base station indicates to the user equipment that at least one of one or more models in the second group is activated.

45. A wireless communication device, comprising at least one processor and a memory, wherein, The at least one processor is configured to cause the device to perform the method of any one of claims 1 to 44.

46. ​​A computer program product comprising a computer-readable program medium including code stored thereon, the code, when executed by a processor, causing the processor to perform the method of any one of claims 1 to 44.

Images