Channel state information (CSI) prediction report processing
By prioritizing CSI reports and grouping them based on available CPU resources, wireless devices effectively generate predicted CSI values, addressing resource constraints and enhancing communication performance.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- GOOGLE LLC
- Filing Date
- 2024-11-08
- Publication Date
- 2026-05-15
AI Technical Summary
Wireless communication devices face challenges in efficiently generating predicted channel state information (CSI) reports due to limited processing and storage resources, leading to inaccurate or discarded CSI values when multiple prediction tasks are required.
A method for prioritizing CSI reports based on available CPU resources, using machine learning models to generate predicted CSI values for high-priority reports while discarding or reporting previous values for low-priority reports, and grouping CSI reports to reduce resource usage.
Enables efficient generation of predicted CSI reports with accurate values, optimizing resource utilization and improving communication quality by adapting to changing conditions.
Smart Images

Figure CN2024130965_15052026_PF_FP_ABST
Abstract
Description
CHANNEL STATE INFORMATION (CSI) PREDICTION REPORT PROCESSINGTECHNICAL FIELD
[0001] This disclosure relates generally to wireless communication and some aspects relate to a CSI report that includes predicted CSI metrics based on machine learning prediction.BACKGROUND
[0002] This background description is provided for the purpose of generally presenting the context of the disclosure. Work of the presently named inventors, to the extent it is described in this background section, as well as aspects of the description that may not otherwise qualify as prior art at the time of filing, are neither expressly nor impliedly admitted as prior art against the present disclosure.
[0003] A network entity and a UE cooperate to determine appropriate operating parameters for communication between the network entity and the UE. For example, the network entity provides a downlink reference signal, such as a channel state information reference signal (CSI-RS) or synchronization signal block (SSB) , to the UE. The UE measures various characteristics of the reference signal and provides channel state information (CSI) feedback to the network entity based on the measured characteristics. The network entity uses the CSI feedback to adjust transmission parameters for communications with the UE, for example, to optimize multiple input multiple output (MIMO) and beamforming operation.
[0004] BRIEF SUMMARY
[0005] The systems, methods, and apparatuses of this disclosure each have several innovative aspects, no single one of which is solely responsible for the desirable attributes disclosed herein.
[0006] One innovative aspect of the subject matter described in this disclosure can be implemented as a method for wireless communication by a wireless device. The method includes a user equipment (UE) receiving, from a network entity, a configuration that includes a plurality of channel state information (CSI) report configurations for a plurality of CSI reports, the plurality of CSI reports including one or more predictive CSI reports. The method includes the UE generating a predictive CSI report of the one or more predictive CSI reports based on a priority for the predictive CSI report from among the plurality of CSI reports and a number of CSI processing units (CPUs) available for one or more prediction tasks associated with the predictive CSI report. The method includes the UE transmitting, to the network entity, the predictive CSI report.
[0007] Another innovative aspect of the subject matter described in this disclosure can be implemented as a method for wireless communication by a wireless device. The method includes a network entity transmitting, to a UE, a configuration that includes a plurality of CSI report configurations for a plurality of CSI reports, the plurality of CSI reports including one or more predictive CSI reports, and a prediction function priority for a prediction function used to predict a CSI value for the one or more predictive CSI reports. The method includes the network entity transmitting a plurality of downlink reference signals. The method includes the network entity receiving, from the UE, at least one predictive CSI report including the predicted CSI value based on the plurality of downlink reference signals.
[0008] Another innovative aspect of the subject matter described in this disclosure can be implemented as an apparatus that includes a communication unit and a processing system configured to control the communication unit to implement any one of the above-referenced methods.
[0009] Details of one or more implementations of the subject matter described in this disclosure are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages will become apparent from the description, the drawings, and the claims.BRIEF DESCRIPTION OF THE DRAWINGS
[0010] Like reference numbers and designations in the various drawings indicate like elements. Note that the relative dimensions of the figures may not be drawn to scale. To easily identify the discussion of any particular element or act, the most significant digit or digits in a reference number refer to the figure number in which that element is first introduced.
[0011] FIG. 1 is a diagram illustrating an example wireless communication system including a user equipment (UE) communicating with a network entity.
[0012] FIG. 2 is a timing diagram illustrating an example channel state information (CSI) report having predicted values.
[0013] FIG. 3 is a communications flow diagram illustrating example operations of a network entity and a UE for coordinating a CSI report having predicted CSI values.
[0014] FIG. 4 is a flow chart diagram illustrating example operations of a UE for generating CSI reports.
[0015] FIG. 5 is a block diagram illustrating example UE capability report information.
[0016] FIG. 6 is a block diagram illustrating an example CSI report configuration.
[0017] FIG. 7 is a block diagram illustrating an example workflow through a machine learning (ML) model.
[0018] FIG. 8 is a block diagram illustrating prompt tuning of a ML model.
[0019] FIG. 9 is a block diagram illustrating an example workflow for a multi-task prediction module.
[0020] FIG. 10 is a block diagram illustrating tokenization options.
[0021] FIG. 11 is a block diagram illustrating differentiating prediction tasks in different time scales by positional encoding.
[0022] FIG. 12 is a flow chart diagram illustrating example UE operations for CSI reporting where CSI reports may include predicted CSI values.
[0023] FIG. 13 is a flow chart diagram illustrating example network entity operations for configuring and receiving CSI reports that may include predicted CSI values.
[0024] FIG. 14 is a flow chart diagram illustrating example UE operations for CSI reporting where CSI reports may include predicted CSI values.
[0025] FIG. 15 is a flow chart diagram illustrating example network entity operations for configuring and receiving CSI reports that may include predicted CSI values.
[0026] FIG. 16 is a block diagram illustrating example configurations of a network entity and a user equipment.DETAILED DESCRIPTION
[0027] The following description is directed to certain implementations for the purpose of describing innovative aspects of this disclosure. However, a person having ordinary skill in the art will readily recognize that the teachings herein can be applied in a multitude of different ways. Some of the examples in this disclosure are based on wireless communication according to the 3rd Generation Partnership Project (3GPP) wireless standards, such as ambient internet-of-things (A-IoT) , the 4th generation (4G) Long Term Evolution (LTE) and 5th generation (5G) New Radio (NR) standards. However, the described techniques can be implemented in any device, system, or network that is capable of transmitting and receiving radio frequency signals according to any of the wireless communication standards, including any of the Institute of Electrical and Electronics Engineers (IEEE) 802.11 or 802.16 wireless standards, or other known signals that are used to communicate within a wireless, cellular, or IoT network, such as a system utilizing 4G, 5G, 6th generation (6G) , ZigBee, Bluetooth, WiFi, or future radio technology.
[0028] As noted above, during a communications session between a network entity and a user equipment (UE) , the UE can measure various characteristics of a downlink (DL) reference signal (RS) from the network entity. Example DL-RSs include as a channel state information reference signal (CSI-RS) or synchronization signal block (SSB) . The UE provides channel state information (CSI) feedback to the network entity based on the measured characteristics. The network entity uses the CSI feedback to adjust transmission parameters for communications with the UE, for example, to optimize multiple input multiple output (MIMO) and beamforming operation.
[0029] The network entity can configure the UE to report the DL-RS feedback to the network entity in one or more CSI reports based on CSI report configurations. A CSI report can include values or indicators for CSI metrics based on its CSI report configuration. Such metrics may include one or more of channel quality indicator (CQI) , rank indicator (RI) ; reference signal received power (RSRP) , reference signal received quality (RSRQ) , received signal strength indicator (RSSI) , signal to noise (SNR) , signal to interference noise ratio (SINR) at the layer 1 (L1) or layer 3 (L3) level. In some aspects, the network entity may configure a CSI report that includes predicted values or indicators for CSI metrics that are based on the output of machine learning (ML) models. The network entity and the UE can use the predicted CSI values to improve communications quality and robustness. For example, the network entity and the UE can anticipate changing conditions of the communications environment and can adapt to these changing conditions before communication between the network entity and the UE is degraded.
[0030] The UE may generate predicted CSI values based on the output of one or more ML models. Predicting values for CSI metrics can tax both processing resources and storage resources of a UE. In some situations, the UE may not have enough processing resources to predict all of the CSI values for CSI reports configured by the network entity. In such situations, the UE may discard CSI reports or may report previously measured or predicted values, which may not be accurate or applicable to current communications conditions. The UE may not have enough storage resources to store multiple ML models that may be used for predicting different CSI values. In some aspects, the UE may implement an ML model capable of performing multiple prediction tasks depending on processing and storage constraints. The processor and storage resource constraints at the UE can lead to uncertainty about the operation of ML model (s) , the accuracy of predicted CSI values, and which CSI reports to generate.
[0031] Various techniques of the disclosure relate to prioritizing CSI reports when the UE does not have available processing resources to perform all prediction tasks needed for a configuration of one or more CSI reports that include predicted CSI values. As used herein, a legacy CSI report will refer to a CSI report that does not include any predicted CSI values, for example, a CSI report that complies with existing specifications. As used herein, a predictive CSI report refers to a CSI report that includes at least one CSI value that is predicted, for example, using ML techniques. As used herein, references to a CSI report (e.g., without a legacy or predictive label) may refer to either a legacy CSI report or a predictive CSI report.
[0032] A network entity transmits a configuration to a UE that configures one or more CSI reports. The one or more CSI reports include at least one predictive CSI report. During operation, the UE receives DL-RSs. The UE measures various CSI metrics such as channel quality indicator (CQI) , rank indicator (RI) ; reference signal received power (RSRP) , reference signal received quality (RSRQ) , received signal strength indicator (RSSI) , signal to noise (SNR) , signal to interference noise ratio (SINR) at the layer 1 (L1) or layer 3 (L3) level. In some aspects, the UE may be configured to use historical values and / or current values of the CSI metrics to predict a future value of a CSI metric. For example, the UE can provide the historical and / or current values of one or more CSI metrics as input to an ML model that has been trained to provide a one or more predicted CSI metrics based on the input.
[0033] The UE uses CSI processing units (CPUs) to perform operations related to gathering, processing, and reporting CSI. These operations may include using an ML model to predict CSI values. During operation of the UE, the UE uses available CPUs to perform CSI related operations. CPUs that are not busy with predictive tasks may be referred to as available CPUs, unoccupied CPUs, or similar terms. CPUs that are busy with predictive tasks may be referred to as unavailable CPUs, occupied CPUs, or similar terms. In some situations, there may not be enough available CPUs to predict CSI values for one or more predictive CSI reports. In such situations, the UE prioritizes the CSI reports based on CPU rules. The CPU rules may determine a priority for a predictive CSI report based on the prediction tasks required for the predictive CSI report, whether the CSI report contains measured CSI values in addition to predictive values, or time stamps associated with the prediction, among other factors. The UE may allocate available CPUs for processing CSI reports based on the priorities associated with the CSI reports. For example, the UE may allocate CPUs for processing CSI reports in the order of CSI report priority. After all available CPUs (or all CPUs configured for generating predictive CSI reports) have been allocated, lower priority CSI reports may not be processed and / or reported. In such cases, the UE may drop (e.g., discard) the lower priority CSI report or may report previous values for the CSI metrics in the CSI report.
[0034] In some aspects, the UE may support groups of predictive CSI reports. The UE may be able to process a group of predictive CSI reports using fewer CPUs than processing each predictive CSI report individually. In some aspects, a group of predictive CSI reports may be processed using a single ML model rather than different ML models for different predictive CSI reports. Processing predictive CSI reports using a single ML model may use fewer storage resources on the UE. For example, a single ML model may be trained for multiple prediction tasks. A task sequence may be prepended or appended to tokens supplied as input to the ML model to guide the ML model to perform a particular prediction task for a prediction.
[0035] One potential technical advantage associated with the techniques of the disclosure is that a UE may be configured to provide predictive CSI reports having predicted CSI values for metrics that are considered to be important by the network while dropping predictive CSI reports having less important metrics. The prioritization of predictive CSI reports provides an efficient way for a UE to provide predictive CSI reports when there are not enough CPUs available to process all requested CSI reports.
[0036] FIG. 1 is a diagram illustrating an example wireless system 100 including a user equipment (UE) 102 communicating with a network entity 104. Although illustrated as a smartphone in FIG. 1, the UE 102 may be implemented as any suitable computing or electronic device, such as a mobile communication device, a modem, cellular phone, gaming device, navigation device, media device, laptop computer, desktop computer, tablet computer, smart appliance, vehicle-based communication system, an Internet-of-things (IoT) device (e.g., sensor node, controller / actuator node, combination thereof) , and the like. The UE 102 may communicate with network entity 104 using wireless links (not shown in FIG. 1) , which may be implemented as any suitable type of wireless link. The wireless links may include one or more wireless links (e.g., radio links) or bearers implemented using any suitable communication protocol or standard, or combination of communication protocols or standards, such as 3GPP LTE, 5G NR, and so forth. Multiple wireless links, referred to as “component carriers” (CCs) may be aggregated in a carrier aggregation to provide a higher data rate for communication between the UE 102 and the network entity 104.
[0037] As an example, the network entity 104 may be a base station, an Evolved Universal Terrestrial Radio Access Network Node B (E-UTRAN Node B) , evolved Node B (eNodeB or eNB) , Next Generation Node B (gNodeB or gNB) , Next Generation E-UTRAN Node B (ng-eNB) , access point, radio head or the like. The network entity 104 may be implemented in a macrocell, microcell, small cell, picocell, or the like, or any combination thereof. The network entity 104 may be configured to use multiple-input-multiple-output (MIMO) communication to exchange wireless signals with the UE 102.
[0038] The network entity 104 supports wireless communication with one or more UEs, such as the UE 102, via radio frequency (RF) signaling using one or more applicable radio access technologies (RATs) as specified by one or more communications protocols or standards. The network entity 104 may employ any of a variety of RATs, such as operating as a NodeB (or base transceiver station (BTS) ) for a Universal Mobile Telecommunications System (UMTS) RAT (also known as “3G” ) , operating as an eNB for a 3GPP LTE RAT, operating as a gNB for a 3GPP 5G NR RAT, and the like.
[0039] The network entity 104 may be part of a radio access network (RAN) , for example, an Evolved Universal Terrestrial Radio Access Network, E-UTRAN, 5G NR RAN, or NR RAN. The network entity 104 may be connected to a Core Network 110. For example, the network entity 104 may connect to the Core Network 110 through an NG2 interface for control-plane signaling and use an NG3 interface for user-plane data communications when connecting to a 5G core network or use an Si interface for control-plane signaling and user-plane data communications when connecting to an Evolved Packet Core (EPC) network. The network entity 104 may communicate using an Xn Application Protocol (XnAP) through an Xn interface or using an X2 Application Protocol (X2AP) through an X2 interface to exchange user-plane and control-plane data. A UE (e.g., UE 102) may connect, via the Core Network 110, to one or more wide area networks (WANs, e.g., WAN 112) or other packet data networks (PDNs) , such as the Internet.
[0040] In some aspects, the functionality, and thus the hardware components, of a network entity such as network entity 104 may be distributed across multiple network nodes or devices and may be distributed in a manner to perform the functions described herein. As one example, the functionality of a network entity (e.g., network entity 104) may be distributed across a radio unit (RU) , a distributed unit (DU) , or a central unit (CU) .
[0041] Communications between a network entity and a UE utilize an uplink (UL) transmission path for transmission path for RF transmissions from the UE to the network entity and a downlink (DL) transmission path for RF transmissions from the network entity to the UE. For example, as shown in FIG. 1, the UE 102 utilizes UL transmission path 116 for RF transmissions from the UE 102 to the network entity 104 and DL transmission path 118 for RF transmissions from the network entity 104 to the UE 102. In the context of the UL transmission path 116, the UE 102 serves as the data sending device and the network entity 104 serves as the data receiving device, whereas in the context of the DL transmission path 118, the network entity 104 serves as the data sending device and the UE 102 serves as the data receiving device. UL transmission path 116 and DL transmission path 118 may utilize multiple communication channels for signal transmission. The multiple channels may each have different purposes. The network entity 104 and the UE 102 may be configured to use MIMO communication in which multiple beams 114 are used to exchange wireless communication signals with UE 102.
[0042] UL transmission path 116 may include a physical uplink shared channel (PUSCH) , a physical uplink control channel (PUCCH) , and a physical random access channel (PRACH) . The PUSCH is used for the transmission of user data, such as voice data, video data, or text message data from the UE 102 to the network entity 104. Additionally, the PUSCH may be used to transmit control information (e.g., uplink control information (UCI) ) . The PUSCH may be shared by multiple UEs. The PUCCH is used for transmitting control information (e.g., UCI) from the UE to the network, such as channel quality feedback, scheduling requests, and acknowledgments. The PRACH is used for random access in the uplink direction, enabling the UE to access the system.
[0043] DL transmission path 118 may include one or more of a physical downlink shared channel (PDSCH) , a physical downlink control channel (PDCCH) , a physical broadcast channel (PBCH) , or a paging channel. The PDSCH is used for transmission of user data from the network entity 104 to the UE 102. The PDSCH may be shared by multiple UEs. As with the PUSCH, the data may be any type of information, such as voice data, video data, or text message data. The paging channel is used to notify a UE that there is incoming traffic for it from a network entity.
[0044] The network entity 104 can configure a Channel State Information (CSI) report configuration for the UE 102 to use to report the CSI. For a CSI report, the UE 102 may report at least one of rank indicator (RI) , precoder matrix indicator (PMI) , channel quality indicator (CQI) , and layer indicator (LI) . RI and PMI are used to indicate the digital precoder; CQI is used to indicate the signal-to-interference plus noise (SINR) status to assist the network entity in determining the modulation and coding scheme (MCS) ; and LI is used to identify the strongest layer for the reported precoder indicated by RI and PMI. The network entity 104 can also configure a CSI report configuration for the UE 102 to use to report the layer 1 reference signal received power (L1-RSRP) or layer 1 signal-to-interference plus noise ratio (L1-SINR) for one or more synchronization signal block (SSB) resources or CSI-RS resources.
[0045] In some aspects, the network entity 104 may configure the UE 102 to report predicted CSI metrics based on ML techniques. During operation, the UE 102 may optionally provide UE capability information 120 to the network entity 104. The UE capability information 120 may indicate the UE's ability to support ML based CSI metric prediction. Depending on the capability reported by the UE 102, the network entity 104 may transmit a CSI report configuration 130 to the UE 102. The CSI report configuration can configure one or more CSI reports that may include legacy CSI reports and predictive CSI reports. The network entity 104 may transmit one or more reference signals 140 to the UE 102. The reference signals 140 may be signals configured measure (and / or predicting) values of CSI metrics. The UE 102 may measure attributes of the reference signals 140. Based on the attributes of the reference signals 140, the UE 102 may use ML based CSI prediction to generate 150 one or more CSI reports 161 that may include either or both legacy CSI reports and predictive CSI reports. The UE may transmit 160 the CSI reports to the network entity 104.
[0046] The UE 102 uses CSI processing units (CPUs) to generate both legacy CSI reports and predictive CSI reports. In some cases, there may not be enough CPUs available to process a CSI report. For example, ML based prediction may occupy multiple CPUs due to the processing capability needed for ML base prediction. If there are not enough available CPUs for generating multiple CSI reports, the UE 102 prioritizes the CSI reports and in some aspects, may discard low priority CSI reports or report previously measured or predicted values for low priority CSI reports. Further details on such prioritization are provided below with respect to FIG. 2 through FIG. 15
[0047] FIG. 2 is a timing diagram 200 illustrating an example of generating a CSI report having predicted values. The network entity 104 may configure the UE 102 with one or more CSI-RSs, some or all of which the UE 102 may use for channel measurements for use in predicting CSI metrics.
[0048] In the example of FIG. 2, at time t1, the UE 102 receives DL-RS 240A from network entity 104. At time t2, the UE 102 receives DL-RS 240B from network entity 104. At time t3, the UE 102 receives DL-RS 240C from network entity 104. DL-RSs 240A, 240B, and 240C may be instances of DL-RS (s) 140 of FIG. 1.
[0049] The UE 102 may obtain, through measurements of DL-RSs 240A-240C, values of one or more metrics associated with the DL-RSs. The metrics obtained through measurement of DL-RSs 240A, 240B, and 240C may be referred to as “measured metrics. ” The UE 102 may provide these measured metrics as input to an ML module 250. The ML module 250 may use machine learning techniques and an ML model 254 to generate predicted CSI values 252 as output of the ML module 250. The ML model 254 may be created (e.g., trained) using different machine learning techniques and / or have coefficients and parameters that are determined via supervised or unsupervised learning. In some aspects, these machine learning techniques may include one or more of neural networks (transformers) , masked language modeling, causal language modeling, self-supervised learning, unsupervised pretraining, and transfer learning, among others. In some aspects, the ML model 254 is trained to provide predicted CSI metrics as output based on input channel metrics. Additionally, or alternatively, the machine learning models may be trained to generate a predicted confidence value 255 (also referred to as a “confidence level” ) reflecting the level of confidence that the ML module 250 associates with the predicted CSI values 252. In some aspects, the confidence value 255 may be the standard deviation of the prediction error, or a probability of the predicted value to be true and / or reliable (within a certain error range) . In some aspects, the confidence value can be derived from assistant information associated with the ML module 250. For example, the confidence value may be derived from an empirical error level based on historical data for the metric. As an example, the historical data may be maintained by a network entity 104 or a vendor of the UE 102 and provided to the UE 102.
[0050] In some aspects, the UE obtains the ML model 254 from a network entity (e.g., network entity 104 of FIG. 1) or core network. In some other aspects, the UE obtains ML model 254 from a vendor of the UE. In some other aspects, the UE obtains ML model 254 from a third party service.
[0051] The UE 102 may generate a predictive CSI report 260 that is based on the predicted CSI values 252. At time t4, the UE 102 transmits the CSI report 260 to the network entity 104.
[0052] FIG. 3 is a communications flow diagram 300 illustrating example operations of a network entity 104 and a UE 102 for coordinating a CSI report having predicted CSI values. Although not illustrated for the sake of illustration clarity, the network entity 104 and / or the UE 102 may implement various acknowledgements for messages illustrated in FIG. 3 to ensure reliable operations for measuring reference signals and providing machine-learning based predictive CSI reports.
[0053] At operation 320, the UE 102 may optionally transmit or report to network entity 104 UE capability information regarding the UE’s capability or support for a machine learning-based CSI reporting. In some aspects, UE 102 may communicate UE capability information to the network entity 104 during an initial communication session setup process between the UE 102 and the network entity 104. In some implementations, the network entity 104 may receive the UE capability information from a core network (e.g., from an access and mobility management function (AMF) of the core network 110 of FIG. 1) . In some implementations, the network entity 104 may receive the UE capability information from another network entity (e.g., a gNB or eNB) . The UE capability information may include supported frequency bands, radio access technologies, maximum transmission power, maximum data rates, and network protocols. In some aspects, the UE 102 may report UE capability information that includes supported configurations for machine learning-based CSI reporting. An example of UE capability information is discussed below with reference to FIG. 5.
[0054] At operation 322, the UE 102 may optionally transmit potential CSI report grouping information to the network entity 104. In some implementations, the UE 102 may support groups of predictive CSI reports. The UE may be able to process a group of predictive CSI reports using fewer CPUs than processing each predictive CSI report individually. The potential report grouping (s) indicate which CSI reports or report type (s) may be grouped together. Grouping CSI reports together for processing may be advantageous because processing a group of CSI reports may use fewer CPUs than the sum of the number of CPUs that would be used for each individual CSI report or report type in the group. The UE 102 may further indicate the required number of CPUs for each group of CSI reports. In some aspects, the required number of CPUs for a CSI report group may be based on the maximum or minimum number of CPUs among the CSI reports in the potential CSI report group. In some aspects, the UE 102 may implement a single ML prediction module to obtain CSI predictions corresponding to the multiple grouped CSI reports. The UE 102 may transmit the report by an RRC message. In some aspects, the UE 102 may transmit the potential CSI report groupings as part of the UE capability information of operation 320. In some aspects, the UE 102 transmits the potential CSI report groupings via UE assistance information (UAI) , via a medium access control (MAC) control element (MAC-CE) , or via UCI.
[0055] At operation 330, depending on the UE capability information that the network entity 104 receives at operation 320, the network entity 104 may transmit to the UE 102, configuration information that configures one or more CSI reports. The CSI reports may include at least one predictive CSI report for reporting predicted CSI metrics based on a machine learning model. In some aspects, the network entity 104 may transmit the configuration via radio resource control (RRC) signaling, e.g., RRCReconfiguration. An example configuration is described below with reference to FIG. 6.
[0056] At operation 324, the UE 102 may optionally transmit a report to the network entity 104 indicating which CSI reports will be grouped together for generating and reporting CSI. In some aspects, the UE 102 may report a required number of CPUs for each group. In some aspects, the UE 102 may report multiple CSI report groups. In some aspects, a CSI report may correspond to at most one CSI report group. In some examples, the UE 102 may transmit the report by an RRC message, for example, UAI, a MAC-CE) or UCI.
[0057] At operation 340, the network entity 104 may transmit the one or more DL-RSs to the UE 102. In some aspects, the network entity 104 may transmit the DL-RSs periodically (P), semi-persistently (SP) , or aperiodically (AP) .
[0058] At block 342, the UE 102 may measure or determine metrics of the one or more DL-RSs that the UE 102 receives at operation 340. For example, the UE 102 may measure channel coefficients, channel quality metrics (e.g., one or more of CQI, RI, RSRP, RSRQ, RSSI, SNR, SINR) .
[0059] At block 350, the UE 102 generates CSI reports. In some aspects, at block 351 the UE 102 calculates priorities associated with requested CSI report (s) or CSI report group (s) . Further details on the determining a priority for a CSI report are discussed below with respect to FIG. 4.
[0060] At block 354, the UE 102 generates CSI reports based on measurements of the DL-RSs that the UE 102 receives at operation 340. When generating a predictive CSI report, the UE 102 provides the current measurements of the DL-RSs as input to an ML model and receives one or more predicted CSI metrics as output from the ML model. In some aspects, the UE 102 may provide both current and past measurements (e.g., a time series of measurements) to the ML model and receives one or more predicted CSI metrics as output. In some aspects, the predicted metrics may correspond to a channel quality or a beam quality at a future time (e.g., time-domain prediction) . In some aspects, the predicted CSI metrics may correspond to a channel quality or beam quality at a current time for a channel or beam where the UE 102 has not been performing measurements (e.g., spatial-domain prediction) .
[0061] If there are not enough available (e.g., unoccupied) CPUs to perform prediction tasks for all of the CSI reports, the UE 102 allocates available CPUs to CSI reports (both legacy and predictive) in the order of the priority associated with the CSI report. For example, the UE 102 allocates CPUs starting with the CSI report with the highest priority and proceeds to allocate CPUs in order of CSI report priority. When the UE 102 no longer has any available CPUs, the UE 102 may omit performing or refrain from performing the prediction tasks for remaining CSI reports (e.g., lower priority CSI reports) .
[0062] At operation 360, the UE 102 transmits the generated CSI reports (legacy CSI reports and / or predictive CSI reports) to the network entity 104.
[0063] At operation 390, the network entity 104 schedules subsequent transmissions and / or updates communication configurations based on the CSI report.
[0064] As an example of the above communication flow, a first report configuration (configuration 1) configures the UE to measure RS associated with setA beams, and predict RSRP of RS associated with setB beams at time t1. A second report configuration (configuration 2) configures the UE to measure RS associated with setA’ beams, and predict RSRP of RS associated with setB beams at time t2. When running the prediction associated with each report configuration above, the required number of CPU is 2. When both report configurations are activated, the UE may implement a single ML module which takes the inputs from measuring RS associated with Set A and A’ , and predicts the RSRPs in both RS associated with set B and B’ . The required number of CPU is 3. In this case, UE may report a CSI report group consisting of report configuration 1 and report configuration 2, and further indicate the required number of CPUs is 3 for the CSI report group.
[0065] In some aspects, a CSI report 1 which contains measured CSI may be grouped with a CSI report 2 which contains predicted CSI, e.g., when the CSI reports 1 and 2 share the same CSI measurement resources.
[0066] FIG. 4 is a flow chart diagram 450 illustrating operations of a UE for generating CSI reports. In general, the operations of FIG. 4 process CSI reports and CSI report groups in order of their respective priorities. As an example, a UE may be configured with N CPUs for generating CSI reports. If L CPUs are in use (e.g., “occupied” ) for calculation of CSI reports or CSI report groups in a given orthogonal frequency-division multiplexing (OFDM) symbol, the UE has (N-L) available CPUs. In some aspects, the occupied symbols (e.g., time-domain symbols allocated for transmission) of a CSI report group is the union of the occupied symbols corresponding to each component CSI report in the group. As an example, if X CSI reports or CSI report groups start occupying their respective CPUs on the same OFDM symbol, and the total required number of CPUs required for processing prediction tasks associated with a CSI report or CSI report group exceeds the total number of available CPUs, the UE may assign the CPUs to the X CSI reports or CSI report groups with higher priority among the X CSI reports or CSI report groups, and not update certain CSI reports or CSI report groups with lower priority.
[0067] At block 451, and as discussed above with respect to FIG. 3, block 351, the UE 102 calculates a priority associated with a CSI report or one or more prediction tasks for a CSI report. In some aspects, for legacy CSI reports, the UE may use existing prioritization heuristics to calculate a priority for the legacy CSI report. In some aspects, for predictive CSI reports, the UE may use one or more of the following to calculate a priority for a predictive CSI report:
[0068] ● A predefined rule which is based on one or more of the following factors:
[0069] ○ Whether the CSI report contains measured CSI
[0070] ○ Time stamp associated with the prediction
[0071] ■ Lower priority for longer-term prediction
[0072] ○ An ML function / module used for the prediction (e.g., a model priority ID, function ID, or module ID)
[0073] ○ Empirical error for the prediction (e.g., the average error based on the past test / validation)
[0074] ○ CC ID associated with the CSI report
[0075] ○ CC ID associated with the CSI prediction
[0076] ○ Report configuration ID for the CSI report
[0077] ○ Types of the CSI report: aperiodic, periodic, semi-persistent, event triggered
[0078] ○ Report quantity for the CSI report (e.g., Whether the CSI report contains L1 RSRP or L1 SINR (either predicted or measured))
[0079] ○ Type of ML functionalities (e.g., spatial domain and / or temporal beam prediction, CSI compression and / or prediction, AI / ML assisted or directed positioning, L3 prediction (e.g., L3-RSRP / L3-SINR / RSRQ prediction) and / or radio link performance prediction)
[0080] ● A UE indication (e.g., indication of priorities among prediction tasks)
[0081] ● A NW indication (e.g., indication of priorities among prediction tasks)
[0082] ● A predefined rule specified in a standards document.
[0083] As discussed above, the UE may group CSI reports together to form a CSI report group. In some aspects, the UE may calculate a priority for a CSI report group as follows:
[0084] ● A predefined rule which is based on one or more of the following factors:
[0085] ○ The priority of each component CSI report
[0086] ■ E.g., the highest priority in the group
[0087] ○ The number of component CSI reports in the group
[0088] ○ One or more properties of the component CSI reports
[0089] ■ Whether the CSI report contains measured CSI
[0090] ■ A time stamp associated with the prediction (e.g., a lower priority may be assigned for a longer-term prediction)
[0091] ■ An ML function, model, or module used for the prediction (e.g., a model priority ID, function ID, or module ID)
[0092] ■ Empirical error for the prediction (e.g., the average error based on the past test / validation)
[0093] ■ CC ID associated with the CSI report
[0094] ■ CC ID associated with the CSI prediction
[0095] ■ Report configuration ID for the CSI report
[0096] ■ Types of the CSI report: aperiodic, periodic, semi-persistent, event triggered
[0097] ■ Report quantity (e.g., whether the CSI report contains L1 RSRP or L1 SINR (either predicted or measured))
[0098] ■ Type of ML functionalities (e.g., spatial domain and / or temporal beam prediction, CSI compression and / or prediction, ML assisted or directed positioning, L3 prediction (e.g., L3-RSRP / L3-SINR / RSRQ prediction) and / or radio link performance prediction)
[0099] ● A UE indication
[0100] ● A network entity indication
[0101] In some implementations, the UE calculates a priority for a CSI report based on the formula: PriiCSI (y, k, c, s, n, P, t, F) =2×Ncells×Ms×y+Ncells×k+Ms×c+s+V×P+Z×t+F×W
[0102] Formula 1
[0103] Where:
[0104] y=0 for aperiodic CSI reports to be carried on PUSCH,
[0105] y=1 for semi-persistent CSI reports to be carried on PUSCH,
[0106] y=2 for semi-persistent CSI reports to be carried on PUCCH,
[0107] y=3 for periodic CSI reports to be carried on PUCCH,
[0108] k=0 for CSI reports carrying L1-RSRP or L1-SINR,
[0109] k=1 for CSI reports not carrying L1-RSRP or L1-SINR,
[0110] c is the serving cell index and Ncells is the value of the higher layer parameter maxNrofServingCells,
[0111] s is the reportConfigID and Ms the value of the higher layer parameter maxNrofCSI-ReportConfigurations,
[0112] P is a variable indicating whether the CSI report contains only predicted CSI: P=1 when the CSI report contains only predicted CSI, and P=0 when the CSI report contains at least one measured CSI,
[0113] V is a fixed value defined in standards or signaled by NW. For instance, the value of V is chosen such that the CSI report carrying only predicted CSI always has a higher priority value (lower priority) than the one which carries at least one measured CSI.
[0114] t is a time stamp associated with the predicted CSI, if any, contained in the CSI report: when P=0, t=0; when P=1, and the CSI report contains multiple predicted CSI, t is determined by the predicted CSI associated with the smallest time stamp ahead in the future.
[0115] Z is a fixed value defined in standards or signaled by the network entity,
[0116] F is a priority order of the prediction function associated with the predicted CSI, if any, contained in the CSI report, where F=0 when P=0; and
[0117] W is a fixed value defined in standards or signaled by the network entity.
[0118] In some aspects, the UE assigns a CSI report that is associated with at least one measured CSI to have a higher priority than any CSI report that is associated with only predicted CSIs. For example, when two CSI reports are associated with the same CSI measurement resources where CSI report 1 contains at least one measured CSI, and CSI report 2 contains only predicted CSI, CSI report 1 always has a higher priority than CSI report 2.
[0119] Block 452 is the top of a loop that allocates CPUs to CSI reports or CSI report groups in the order of a priority associated with the CSI report or CSI report group. For example, the UE may select the CSI report or CSI report group with the highest priority for CPU allocation. At block 452, the UE determines a number of CPUs that are currently available for use for generating CSI metrics.
[0120] At decision block 453, the UE determines if there are enough available CPUs for the number of CPUs required to process the CSI report (e.g., to perform prediction tasks for the CSI report) . If there are enough available CPUs, ( "Yes" branch of decision block 453) , the UE proceeds to block 454 where the UE allocates CPUs to the CSI report or CSI report group for generating the CSI report or CSI reports in a CSI report group.
[0121] If there are not enough available CPUs for the number of CPU for the CSI report ( "No" branch of decision block 453) , the UE proceeds to block 456, where in some aspects, the UE drops (e.g., discards) the CSI report or report group. Additionally, or alternatively, the UE may indicate the reported channel metric in the CSI report is not updated, e.g., due to the limitation of the CPU availability. In some aspects, the UE may use a reserved bit sequence or dedicated bits or bit fields to indicate that the channel metric has not been updated. In some aspects, the UE may report the previous value for the channel metric that has not been updated.
[0122] At decision block 455, the UE determines if there are any remaining CSI reports requiring CPU allocation. If there are remaining CSI reports or report groups to be allocated CPUs ( "Yes branch of decision block 455) , the method returns to block 452 to process the CSI report or report group having the highest priority among the remaining CSI reports or report groups. If there are no remaining CSI reports or report groups to be allocated CPUs ( "No" branch of decision block 455) , the method ends.
[0123] FIG. 5 is a block diagram illustrating example UE capability information 520. In some aspects, the example UE capability information 520 may include one or more of the indicators discussed below.
[0124] The UE capability information can include an indicator 521 that indicates whether UE 102 supports reporting predicted CSI metrics in a CSI report.
[0125] The UE capability information 520 can include the supported prediction tasks 522A. In some aspects, UE may indicate to the network entity the supported types of CSI prediction tasks. The UE may further indicate the number of CPUs 522B for each CSI prediction task or task type. The network entity may configure, activate, and or trigger a CSI report associated with one or more CSI prediction tasks based on the supported prediction tasks 522A.
[0126] The UE capability information 520 can include the number of supported simultaneous CSI calculations 522C. In some aspects, the UE indicates the number of simultaneous CSI calculations in a component carrier (simultaneousCSI-ReportsPerCC) and a number of simultaneous CSI calculations across all component carriers (simultaneousCSI-ReportsAllCC) . If a UE supports N simultaneous CSI calculations the UE is said to have N CPUs for processing CSI reports.
[0127] The UE capability information 520 can include an indicator indicating the number of CPUs that may be used for legacy CSI reports 522D (per CC or in all CCs) .
[0128] The UE capability information 520 can include an indicator indicating the number of CPUs that may be used for CSI reports containing at least one predicted CSI 522E.
[0129] In some aspects, CSI reports associated with CSI prediction may be of different types, for example, based on the report quantities associated with the CSI report, e.g., predicted-RSRP, predicted-RSRQ, predicted-SINR (in L1 or L3) , predicted channel response, predicted RI, predicted CQI, predicted precoder matrix, predicted-beam ID or “none. ” For each CSI report or CSI report type 523, the UE may indicate in the UE capability information 520 the required number of CPUs 524 for the CSI report or report type. The required number of CPUs may be based on the UE's computation capability. In some aspects, the UE may update the required number of CPUs. For example, the UE may update the Required Number of CPUs 524 for a CSI report or report type 523 when there is a change in power availability, hardware configuration, or hardware availability. In some aspects, the required number of CPUs for each CSI report or CSI report type may be defined in standards.
[0130] In some aspects, a predictive CSI report may contain the CSI predictions from multiple CSI prediction tasks. The Required Number of CPUs 524 may be computed based on the required CPU numbers of each involved CSI prediction tasks. For instance, if each of the involved CSI prediction tasks require a dedicated ML module, then the CPU number of the CSI report may be the sum of the CPU numbers from each of prediction tasks associated with the CSI report.
[0131] Different numbers of CPUs may be defined for different categories of UE capabilities. In some aspects, the CSI capability information 520 may indicate an inference configuration 522F for one or more ML functionalities, e.g., spatial domain and / or temporal beam prediction, CSI compression and / or prediction, ML assisted or directed positioning, L3 prediction (e.g., L3-RSRP / L3-SINR / RSRQ prediction) and / or radio link performance prediction.
[0132] In some implementations, the UE may report the number of CPUs for different types of inference configuration, different ML functionalities, and / or different ML models by an RRC message, such as UE capability, UAI, MAC-CE or UCI on PUSCH or PUCCH. In some implementations, for an inference configuration for multiple ML functionalities, e.g., joint CSI compression and prediction, or joint beam prediction in time and spatial domain, the UE may report a separate number of CPUs. Alternatively, the number of CPUs for such inference configuration may be based on the maximum, minimum or total number of CPUs for the corresponding ML functionalities.
[0133] In some aspects, UE may have multiple binaries of ML models or modules to perform the same prediction tasks. Different binaries may correspond to different numbers of required CPU numbers. For example, an ML model or module may have a deeper neural network which requires more CPUs but has better accuracy than the ML models or modules available to the UE. In some aspects, UE may include multiple CPU values 522G for the same prediction task (e.g., ML functions) in the UE capability information 520. When performing the prediction task, the UE may down-select the binary that is actually used for the CSI report. UE may indicate the down-selected binary along with the CSI value predicted by the binary in a CSI report. In some aspects, UE may update the CPU number 522G for a prediction task, when it chooses to use a binary requiring a different CPU number.
[0134] FIG. 6 is a block diagram illustrating an example configuration 630. In some aspects, the parameters and / or information in the example CSI report configuration 631 may include one or more of the following:
[0135] ● Resource set 632A for the UE 102 to use to send the CSI report. Resource set 632A may include one or more PUCCH or PUSCH resources.
[0136] ● Reported measured value (s) 632B indicating the CSI metrics to be included in the CSI report.
[0137] ● Reported predicted value (s) 632C indicating the CSI metrics for which the UE is to predict values for inclusion in the CSI report.
[0138] ● Time indicator (s) 632D indicating how far into the future the UE is to predict.
[0139] Configuration 630 may include activated CSI reports 633 and / or activated CSI report groups 634 indicating which CSI reports or CSI report groups that the UE should report to the network entity.
[0140] Configuration 630 may include ML model (s) 635 configuring the ML models available for by the UE for predicting CSI values. Configuration 630 may include activated ML model (s) s 636 indicating which of ML model (s) 635 are to be used for predicting CSI values.
[0141] FIG. 7 is a block diagram illustrating an example workflow through a machine learning (ML) model 754.
[0142] The raw input ML model may be the DL-RS metrics 740, which may be referred to as the prompt for the ML model. Each measured metric in the prompt is quantized and then tokenized 752 (e.g., mapped) to a predefined token such that the quantized values are mapped to a sequence of tokens. In some cases, additional tokens may be prepended or appended to the sequence of the tokens, e.g., to indicate the start or end of the prompt.
[0143] Then, the sequence of the tokens is fed into an embedding layer 753. In some examples, linear layers may be used for the embedding process. The embedding layer 753 converts the stream of tokens into dense, low-dimensional vectors that capture relationships between these tokens.
[0144] In some aspects, positional encoding 757 may be added to the output of the embedding layer 753. The positional encoding for each token may be determined by its position in the sequence of the tokens. The values of the positional encoding for different positions may be trainable or determined by a predefined formula.
[0145] The output of the embedding layer 753 (including positional encoding, if any) is fed into the ML model 754. In some aspects, ML model 754 may be a transformer model. The ML model 754 may be trained to predict CSI metrics based on the input DL-RS metrics 740.
[0146] The output of the transformer model can be further fed into an output layer. The output layer 755 generates a predicted CSI value 758. In some implementations, the output layer 755 may be a softmax layer. The softmax layer is a layer used in ML models that converts the output of the ML model into a probability distribution indicating a probability associated with a predicted value.
[0147] FIG. 8 is a block diagram illustrating prompt tuning of a ML model. In the example shown in FIG. 8, ML model 854 has been trained to receive tokens (e.g., tokens mapped from quantized RS metrics) and to output predicted CSI value (s) . In some aspects, the ML model 854 can be “fine-tuned” with respect to multiple tasks using a process call prompt tuning. In the example of FIG. 8, the ML model 854 is frozen (e.g., the prediction parameters remain unchanged) . Each task is assigned with a sequence of a few trainable soft tokens (referred to as a “soft prompt” ) , whose values are determined during the fine-tune training process. The value of the soft prompt is determined by the training process, and may not correspond to any existing token. After training, different tasks identified during the fine tuning training will have different soft prompts. In the example of FIG. 8, tokens for different instances of task A (e.g., tokens 856A-856B representing different time series of measured RS data) are input to ML model 854. Similarly, tokens for different instances of task B (e.g., tokens 856E-856H) are input to ML model 854. The ML model learns different tasks, and output learned task sequences 857. Learned task sequences 857 include task sequence 857A for tokens 856A-856D and assigns task sequence 857B for tokens 856E-856H.
[0148] The fine-tuned ML model 854 may be provided to the UE. The UE can the select a task sequence 857 based on the task, and prepend the selected task sequence to the input token sequence 856 before being input to the ML model 854. The number of tokens in the task sequence 857 is an order of magnitude smaller than the parameters in the ML model 854, is an efficient means of guiding the ML model when predicting CSI metrics.
[0149] FIG. 9 is a block diagram illustrating a workflow for a multi-task prediction module 950. A multi-task prediction module is a module that includes an ML model that has been trained to handle multiple different prediction tasks. In some aspects, the channel prediction tasks associated with multiple CSI reports may be computed by running the single ML module. In some cases, the required number of CPUs when running multiple prediction tasks using a single ML module may be fewer than the sum of the CPU numbers when running each individual prediction task.
[0150] In the example of FIG. 9, two sets of DL-RS measurements 940A and 940B are each associated with a different prediction task. The ML model 954 has been fine-tuned as described above to recognize different task sequences for different tasks. The different tasks may correspond to predicting different metrics, predicting metrics from different beams, cells, geographic areas, or in different time scales. The multiple tasks may share the same input format; however, each task may require a different set of measurements and / or a different tokenizing approach. Additionally, or alternatively, different tasks may be associated with different report configurations, which include report format, container of the report (e.g., UCI report, MAC-CE, PUCCH or PUSCH) , report resource (e.g., periodicity) , and / or priority of the report. A dedicated task sequence 957 may be incorporated with the module input, to indicate the nature of the task to the ML model 954. The ML model 954 can use the task sequence 957 to identify the task and output the intended prediction. The different task sequences 957 may be provided to the UE by the network entity or they may be predefined.
[0151] As shown in FIG. 9, the two sets of DL-RS measurements 940A and 940B are input to the ML model 954. DL-RS measurements 940A and 940B may share the same input format, but may be prepared for submission to the ML model 954 in different ways. In some aspects, the input data (e.g., DL-RS measurements 940A and 940B) is two-dimensional, e.g., X by T, where T is the length of the sequence and X is the dimension or bit width of each entry in the input. In some aspects, X is the same for all tasks, while T may vary between tasks.
[0152] The preparation methods (e.g., quantization, tokenization) may be different for different tasks. The network entity may indicate the preparation methods for the task or the preparation methods may be predefined. In some aspects, the DL RL measurements 940A and 940B are quantized 951A and 951B respectively. The quantization may use predefined indices. In some aspects, special indices reserved for the case where the RS measurement is out of the quantization range (too low or too high) or when the measurement is unavailable (e.g., skipped due to power saving or scheduling conflicts) .
[0153] The quantized DL-RS measurements 940A and 940B are then tokenized 952A and 952B, producing tokens 956A and 956B, respectively. As discussed above, tokenization maps one or more of the (quantized) measurements to one or more tokens. In some aspects, different tasks may use different tokenization options, as will be further described with reference to FIG. 10.
[0154] The task sequence associated with the request prediction task is prepended to the tokens to indicate the prediction task to the ML model 954. Additionally, or alternatively, the task sequence may be appended to the tokens or inserted at a predefined location. In the example of FIG. 9, task A sequence 957A is prepended to RS A tokens 956A and task B sequence 957B is prepended to RS B tokens 956B.
[0155] The tokens and their associated task sequences are provided as input to the ML model 954. The ML model, guided by the task sequences, outputs predicted CSI values based on the tokens. For example, ML model 954 outputs predicted CSI values 958A based on task sequence 957A and tokens 956A. ML model 954 output predicted CSI values 958B based on task sequence 957B and tokens 956B.
[0156] In some aspects, different DL-RS measurements may have different accuracies. For example, a first UE using ML model 954 may have a measurement accuracy of + / -1 dB, while a second UE the same ML model 954 may have a better accuracy of + / -0.5 dB. In the quantization step 951, the step of the quantization table for the first UE may be 2 dB, while the step for the second UE may be 1 dB. The second UE uses a more refined quantization table with more indices. This may also lead to different tokenization methods, as the first UE and the second UE will have a different number of quantized indices to be mapped to tokens. In some aspects, the UE may report its capability with respect to RS measurement accuracy. The UE may apply different quantization and tokenization methods based on its reported capability.
[0157] FIG. 10 is a block diagram illustrating tokenization options. As discussed above, different tasks (e.g., tasks associated with different task sequences) may require different preparation methods (e.g., different quantization and tokenization) . In the example shown in FIG. 10, a series of DL-RS measurements 1040A -1040C are to be prepared for input to an ML model (e.g., ML model 954 of FIG. 10) . Additionally, the series includes one DL-RS measurement 1041 that is out of range. The DL-RS measurements are quantized 1051 into quantization indices. In the example of FIG. 10, DL-RS measurements 1040A -1040C are quantized into indices ‘011’ , ‘013’ , and '004', respectively. DL-RS measurement 1041 is quantized to an index ‘999’ indicating the DL-RS is out of range. The quantization indices are tokenized 1052 into tokens 1053. In the example of FIG. 10, the bit values of the token associated with DL-RS measurements 1040A are represented by an ‘a’ , the bit values of the token associated with DL-RS measurements 1040B are represented by a ‘b’ , the bit values of the token associated with DL-RS measurements 1040C are represented by an ‘c’ , and the bit values of the token associated with DL-RS measurements 1041 are represented by an ‘x’ .
[0158] In the example of FIG. 10, there are three different options for representing a token stream associated with DL-RS measurements. In option 1008, each token is eight bits wide and the four tokens in this example occupy four bytes. In option 1010, each token is four bits wide and the four tokens in the example occupy two bytes. In the option 1012, each token is two bits wide and the four tokens in the example occupy a single byte.
[0159] FIG. 11 is a block diagram illustrating differentiating prediction tasks in different time scales by positional encoding. The tokenize operation 752, embed layer 753, positional encoding 757, ML model 754, and output layer 755 have been described above with respect to FIG. 7.
[0160] In the example of FIG. 11, a task specific sequence may include positional encoding sequences or indicate the selection of the positional encoding. In this example, task 1 (represented by box 1141A) is to measure RS1 in the previous four slots (e.g., slots at k=1-4) as well as in the current slot (k=5) and predict the RSRP of RS1 in the next slot (e.g., k=6) . Task 2 (represented by box 1141B) is to measure RS1 in the previous four slots (e.g., k=1-4) as well as the current slot (e.g., k=5) and predict the RSRP of RS1 in a future time that is five slots ahead of the current time slot (e.g., k=10) . Different positional encoding entries of slots 1040 may be selected from a predefined dictionary (wherein each entry is indexed by k) for task 1 and task 2 to address the difference in the prediction time intervals.
[0161] In conventional systems, a traditional recurrent neural network (RNN) may be used to predict the RSRP N slots ahead. In such systems, the RNN is run in an auto-regression manner. For example, the RNN uses the measurement from the four previous slots to the current slot (e.g., slot -4 to slot 0) to predict the RSRP in the next slot (e.g., slot 1) , and then uses the predicted RSRP in slot 1 as input to predict RSRP in slot 2, which requires running the RNN N times. In the technique of FIG. 11, a single pass through the ML model can generate the prediction data by selecting the appropriate positional encoding as shown in FIG. 11.
[0162] FIG. 12 is a flow chart diagram illustrating example UE operations 1200 for CSI reporting where CSI reports may include predicted CSI values. The example operations 1200 may be performed, for example, by the UE 102 of FIG. 1 and FIG. 3.
[0163] At block 1220, and as described above with respect to FIG. 3, operation 320, the UE may optionally transmit or report to network entity 104 UE capability information regarding the UE’s capability or support for a machine learning-based CSI reporting. In some aspects, UE may communicate UE capability information to the network entity during an initial communication session setup process between the UE and the network entity. The UE capability information may include supported frequency bands, radio access technologies, maximum transmission power, maximum data rates, and network protocols. In some aspects, the UE may report UE capability information that includes supported configurations for machine learning-based CSI reporting.
[0164] At block 1222, and as described above with respect to FIG. 3, operation 322, the UE may optionally transmit potential CSI report grouping information to the network entity. The potential report grouping (s) indicate which CSI reports or report type (s) may be grouped together. The UE may further indicate the required number of CPUs for each group of CSI reports.
[0165] At block 1230, and as described above with respect to FIG. 3, operation 330, the UE may receive configuration information that configures one or more CSI reports. The CSI reports may include at least one predictive CSI report for reporting predicted CSI metrics based on a machine learning model.
[0166] At block 1224, and as described above with respect to FIG. 3, operation 324, the UE may optionally transmit a report to the network entity indicating which CSI reports will be grouped together for generating and reporting CSI.
[0167] At block 1240, and as described above with respect to FIG. 3, operation 340, the UE may receive the one or more DL-RSs.
[0168] At block 1242, and as described above with respect to FIG. 3, block 342, the UE may measure or determine metrics of the one or more DL-RSs (from block 1240) . For example, the UE may measure channel coefficients, channel quality metrics (e.g., one or more of CQI, RI, RSRP, RSRQ, RSSI, SNR, SINR) .
[0169] At block 1251, and as described above with respect to FIG. 3, block 351 and FIG. 4, block 451, the UE may calculate a priority associated with a CSI report or CSI report group.
[0170] At block 1254, and as described above with respect to FIG. 3, operation 322, the UE may generate CSI reports based on measurements of the DL-RSs (from block 1240) . When generating a predictive CSI report, the UE may provide the current measurements of the DL-RSs as input to an ML model and receives one or more predicted CSI metrics as output from the ML model. In some aspects, the UE may provide both current and past measurements (e.g., a time series of measurements) to the ML model and receives one or more predicted CSI metrics as output.
[0171] At block 1260, and as described above with respect to FIG. 3, operation 322, the UE may transmit the generated CSI reports (legacy CSI reports and / or predictive CSI reports) to the network entity.
[0172] FIG. 13 is a flow chart diagram illustrating example network entity operations 1300 for configuring and receiving CSI reports that may include predicted CSI values. The example operations 1300 may be performed, for example, by the network entity 104 of FIG. 1 and FIG. 3.
[0173] At block 1320, and as described above with respect to FIG. 3, operation 320, the network entity may optionally receive UE capability information regarding a UE’s capability or support for a machine learning-based CSI reporting.
[0174] At block 1322, and as described above with respect to FIG. 3, operation 322, the network entity may optionally receive potential CSI report grouping information from the UE. The potential report grouping (s) indicate which CSI reports or report type (s) may be grouped together, and may include the required number of CPUs for each group of CSI reports.
[0175] At block 1330, and as described above with respect to FIG. 3, operation 330, the network entity may transmit configuration information that configures one or more CSI reports. The CSI reports may include at least one predictive CSI report for reporting predicted CSI metrics based on a machine learning model.
[0176] At block 1324, and as described above with respect to FIG. 3, operation 324, the network entity may optionally receive a report from the UE indicating which CSI reports will be grouped together for generating and reporting CSI.
[0177] At block 1340, and as described above with respect to FIG. 3, operation 340, the network entity may transmit the one or more DL-RSs indicated in the configuration (from block 1330) .
[0178] At block 1360, and as described above with respect to FIG. 3, operation 360, the network entity may receive CSI reports (legacy CSI reports and / or predictive CSI reports) generated by the UE.
[0179] At block 1390, and as described above with respect to FIG. 3, operation 390, the network entity schedules subsequent communications between the UE 102 and the network entity 104 based on the CSI reports.
[0180] FIG. 14 is a flow chart diagram illustrating example UE operations 1400 for CSI reporting where CSI reports may include predicted CSI values. The example operations 1400 may be performed, for example, by the UE 102 of FIG. 1 and FIG. 3.
[0181] In block 1430, and as described above with respect to FIG. 3, operation 330 and FIG. 12, block 1230, the UE receives, from a network entity, a configuration that includes a plurality of CSI report configurations for a plurality of CSI reports, the plurality of CSI reports including one or more predictive CSI reports.
[0182] In block 1450, and as described above with respect to FIG. 3, block 350, the UE generates a predictive CSI report of the one or more predictive CSI reports based on a priority for the predictive CSI report from among the plurality of CSI reports and a number of CPUs available for one or more prediction tasks associated with the predictive CSI report.
[0183] In block 1460, and as described above with respect to FIG. 3, operation 360 and FIG. 12, block 1260, the UE transmits the predictive CSI report to the network entity.
[0184] FIG. 15 is a flow chart diagram illustrating example network entity operations 1500 for configuring and receiving CSI reports that may include predicted CSI values. The example operations 1500 may be performed, for example, by the network entity 104 of FIG. 1 and FIG. 3.
[0185] In block 1530, and as described above with respect to FIG. 3, operation 330 and FIG. 13, block 1330, the network entity transmits a configuration that includes a plurality of CSI report configurations for a plurality of CSI reports, the plurality of CSI reports including one or more predictive CSI reports, and a prediction function priority for a prediction function used to predict a CSI value for the one or more predictive CSI reports.
[0186] In block 1540, and as described above with respect to FIG. 3, operation 340 and FIG. 13, block 1340, network entity transmits DL-RSs.
[0187] In block 1560 and as described above with respect to FIG. 3, operation 360 and FIG. 132, block 1360, the network entity receives at least one predictive CSI report including the predicted CSI value based on the plurality of downlink reference signals.
[0188] FIG. 16 is a block diagram illustrating example configurations of a network entity 1604 and a user equipment 1602. Note that the depicted hardware configurations represent the processing components (e.g., a processing system) and communication components (e.g., a communication unit) of a network entity 1604 (such as the network entity 104 described herein) and a UE 1602 (such as the UE 102 described herein) . The depicted hardware configurations may omit certain components well-understood to be frequently implemented in such electronic devices, such as displays, peripherals, power supplies, and the like.
[0189] The UE 1602 includes antennas 1603A, a radio frequency front end (RF front end) 1603B, and radio-frequency transceivers (e.g., an LTE transceiver 1603D and a 5G NR transceiver 1603C) for communicating with the network entity 1604. The RF front end 1603B includes one or more modems configured for the corresponding RAT (s) employed (for example, Third Generation Partnership Project (3GPP) Fifth Generation New Radio (5G NR) ) , one or more analog-to-digital converters (ADCs) , one or more digital-to-analog converters (DACs) , signal processors, and the like. In the example illustrated in FIG. 16, the RF front end 1603B of the UE 1602 may couple or connect the 5G NR transceiver 1603C to the antennas 1603A to facilitate various types of wireless communication. The RF front end 1603B operates, in effect, as a physical (PHY) transceiver interface to conduct and process signaling between the one or more processor (s) 1603E and antennas 1603A so as to facilitate various types of wireless communication.
[0190] The antennas 1603A of the UE 1602 include an array of multiple antennas that may be tuned to one or more frequency bands associated with a corresponding RAT. The antennas 1603A and the RF front end 1603B are tuned to, and / or be tunable to, one or more frequency bands defined by the 3GPP 5G NR communication standards and implemented by the 5G NR transceiver 1603C. Additionally, the antennas 1603A, the RF front end 1603B, and / or the 5G NR transceiver 1603C can be configured to support beamforming for the transmission and reception of communications with the network entity 1604. By way of example and not limitation, the antennas 1603A and the RF front end 1603B may be implemented for operation in sub-gigahertz bands, sub-6 GHz bands, and / or above 6 GHz bands that are defined by the 3GPP LTE and 5G NR communication standards.
[0191] The UE 1602 also includes processor (s) 1603E and computer-readable storage media (CRM) 1603F. The processor (s) 1603E may include, for example, one or more central processing units, graphics processing units (GPUs) , or other application-specific integrated circuits (ASIC) , and the like. To illustrate, the processor (s) 1603E may include an application processor (AP) utilized by the UE 1602 to execute controller functions, an operating system, or various applications, as well as one or more processors utilized by modems or a baseband processor of the RF front end 1603B. The CRM 1603F may include any suitable memory or storage device such as random-access memory (RAM) , static RAM (SRAM) , dynamic RAM (DRAM) , non-volatile RAM (NVRAM) , read-only memory (ROM) , Flash memory, solid-state drive (SSD) or other mass-storage devices, and the like useable to store one or more sets of executable software instructions and associated data that manipulate the one or more processor (s) 1603E and other components of the UE 1602 to perform the various functions described herein and attributed to the UE 1602. The sets of executable software instructions include, for example, an operating system (OS) and various drivers (not shown) , and various software applications (not shown) , which are executable by processor (s) 1603E to enable user-plane communication, control-plane signaling, and user interaction with the UE 1602.
[0192] The processor (s) 1603E along with other processors of the UE 1602 that are used to implement the techniques described herein may be individually or collectively referred to as a “processing system. ” One or more of RF front end 1603B, 5G NR transceiver 1603C, and LTE transceiver 1603D may be individually or collectively referred to as a “communication unit. ”
[0193] In some aspects, the CRM 1603F may include one or more ML models 1654. The ML models 1654 may be any of ML models 254, 754, 854, or 954 of FIGs. 2, 6, 8, or 9, respectively.
[0194] Turning to the hardware of the network entity 1604, it is noted that although FIG. 16 illustrates an implementation of the network entity 1604 as a single network node (for example, a 5G NR Node B, or “gNB” ) , the functionality, and thus the hardware components, of the network entity 1604 instead may be distributed across multiple network nodes or devices and may be distributed in a manner to perform the functions described herein. As one example, the functionality of network entity 1604 may be distributed across a radio unit (RU) , distributed unit (DU) , or central unit (CU) .
[0195] The network entity 1604 includes antennas 1605A, a radio frequency front end (RF front end) 1605B, and one or more 5G NR transceivers 1605C for communicating with the UE 1602. The RF front end 1605B of the network entity 1604 may couple or connect the 5G NR transceivers 1605C to the antennas 1605A to facilitate various types of wireless communication. Similar to RF front end 1603B, the RF front end 1605B includes one or more modems, one or more ADCs, one or more DACs, and the like. RF front end 1605B receives the one or more RF signals, for example, RF signals from UE 1602, and pre-processes the one or more RF signals to generate data from the RF signals. The data is provided as input to processes and / or applications executing on network entity 1604. This pre-processing may include, for example, power amplification, conversion of band-pass signaling to baseband signaling, initial analog-to-digital conversion, and the like.
[0196] The antennas 1605A of the network entity 1604 may be configured individually and / or as one or more arrays of multiple antennas. The antennas 1605A and the RF front end 1605B may be tuned to, and / or be tunable to, one or more frequency band defined by the 3GPP 5G NR communication standards, and implemented by the 5G NR transceivers 1605C. Additionally, the antennas 1605A, the RF front end 1605B, and the 5G NR transceivers 1605C may be configured to support beamforming, such as Massive-MIMO, for the transmission and reception of communications with the UE 1602.
[0197] The network entity 1604 also includes processor (s) 1605D and computer-readable storage media (CRM) 1605E. The processor (s) 1605D may include, for example, one or more central processing units, graphics processing units (GPUs) , or other application-specific integrated circuits (ASIC) , and the like. To illustrate, the processor (s) 1605D may include an application processor (AP) utilized by the network entity 1604 to execute an operating system and various user-level software applications, as well as one or more processors utilized by modems or a baseband processor of the RF front end 1605B to enable communication with the UE 1602. In at least some aspects, the processor (s) 1605D configures the 5G NR transceiver (s) 1605C for communication with the UE 1602, transmission and reception points (TRPs) , and radio units via fronthaul interface 1607A, as well as communication with a core network. In some aspects, the network entity 1604 includes an inter-network entity interface 1607B, such as an Xn and / or X2 interface, which the processor (s) 1605D configures to exchange user-plane and control-plane data with another network entity, to manage the communication of the network entity 1604 with the UE 1602. The network entity 1604 includes a core network interface 1607C that the processor (s) 1605D configures to exchange user-plane and control-plane data with core network functions and entities.
[0198] The processor (s) 1605D along with other processors of the network entity 1604 that are used to implement the techniques described herein may be individually or collectively referred to as a “processing system. ” One or more of RF front end 1605B, 5G NR transceiver (s) 1605C, fronthaul interface 1607, inter-network entity interface 1607B, and core network interface 1607C may be individually or collectively referred to as a “communication unit. ”
[0199] FIG. 1 through FIG. 16 and the operations described herein are examples meant to aid in understanding example implementations and should not be used to limit the potential implementations or limit the scope of the claims. some implementations may perform additional operations, fewer operations, operations in parallel or in a different order, and some operations differently.
[0200] The following additional considerations may apply to the foregoing and the following discussions.
[0201] Unless defined otherwise, technical and scientific terms used herein have the same meaning as is commonly understood by one of ordinary skill in the art to which this specification belongs. The terms “first, ” “second, ” and the like, as used herein do not denote any order, quantity, or importance, but rather are used to distinguish one element from another. The use of terms “including, ” “comprising” or “having” and variations thereof herein are meant to encompass the items listed thereafter and equivalents thereof as well as additional items. The terms “connected” and “coupled” are not restricted to physical or mechanical connections or couplings and can include electrical connections or couplings, whether direct or indirect. Furthermore, terms “circuit” and “circuitry” and “control unit” may include either a single component or a plurality of components, which are either active and / or passive and are connected or otherwise coupled together to provide the described function. In addition, the term operationally coupled as used herein includes wired coupling, wireless coupling, electrical coupling, magnetic coupling, radio communication, software based communication, or combinations thereof.
[0202] Some or all of the foregoing or the following implementations can be jointly combined or formed to be a new or another one implementation. The foregoing or the following techniques can be used to solve at least (but not limited to) the issue (s) or scenario (s) mentioned in this disclosure. Any two or more than two of the foregoing or the following paragraphs, (sub) -bullets, points, actions, or claims described in each method / technique / implementation may be combined logically, reasonably, and properly to form a specific method. Any sentence, paragraph, (sub) -bullet, point, action, or claim described in each of the foregoing or the following technique (s) / implementation (s) / concept (s) may be implemented independently and separately to form a specific method. Dependency, such as “based on, ” “more specifically, ” “where” or etc., in technique (s) / implementation (s) / concept (s) mentioned in this disclosure is just one possible implementation which would not restrict the specific method.
[0203] Generally speaking, description for one of the above figures can apply to another of the above figures. Examples, implementations and methods described above can be combined, if there is no conflict. An event or block described above can be optional or omitted. For example, an event or block with dashed lines in the figures can be optional. In some implementations, “message” is used and can be replaced by “information element (IE) , ” and vice versa. In some implementations, “IE” is used and can be replaced by “field, ” and vice versa. In some implementations, “configuration” can be replaced by “configurations” or “configuration parameters, ” and vice versa. In some implementations, “some” means “one or more. ” In some implementations, “at least one” means “one or more. ”
[0204] As used herein, the terms “wireless device” , “user device” , “user equipment” , “wireless communication device” , “mobile communication device” , “communication device” , or “mobile device” refer to any one or all of cellular telephones, smartphones, portable computing devices, personal or mobile multi-media players, laptop computers, tablet computers, smartbooks, Internet-of-Things (IoT) devices, palm-top computers, wireless electronic mail receivers, multimedia Internet enabled cellular telephones, wireless gaming controllers, display sub-systems, driver assistance systems, vehicle controllers, vehicle system controllers, vehicle communication system, infotainment systems, vehicle telematics systems or subsystems, vehicle display systems or subsystems, vehicle data controllers, point-of-sale (POS) terminals, health monitoring devices, drones, cameras, media-streaming dongles or another personal media devices, wearable devices such as smartwatches, wireless hotspots, femtocells, broadband routers or other types of routers, and similar electronic devices which include a programmable processor and memory and circuitry configured to perform operations as described herein. Further, the user device may be embedded in an electronic system such as the head unit of a vehicle or an advanced driver assistance system (ADAS) . Still further, the user device can operate as an internet-of-things (IoT) device or a mobile-internet device (MID) . Depending on the type, the user device can include one or more general-purpose processors, a computer-readable memory, a user interface, one or more network interfaces, one or more sensors, etc.
[0205] Certain techniques are described in this disclosure as including logic or a number of components or modules. Modules can be software modules (e.g., code, or machine-readable instructions stored on non-transitory machine-readable medium) or hardware modules. A hardware module is a tangible unit capable of performing certain operations and may be configured or arranged in a certain manner. A hardware module can comprise dedicated circuitry or logic that is permanently configured (e.g., as a special-purpose processor, such as a field programmable gate array (FPGA) or an application-specific integrated circuit (ASIC) , a digital signal processor (DSP) , etc. ) to perform certain operations. A hardware module may also comprise programmable logic or circuitry (e.g., as encompassed within a general-purpose processor or other programmable processor) that is temporarily configured by software to perform certain operations. The decision to implement a hardware module in dedicated and permanently configured circuitry, or in temporarily configured circuitry (e.g., configured by software) may be driven by cost and time considerations.
[0206] When implemented in software, the techniques can be provided as part of the operating system, a library used by multiple applications, a particular software application, etc. The software can be executed by one or more general-purpose processors or one or more special-purpose processors.
[0207] As used herein, the terms “component” and “module” are intended to be broadly construed as hardware, firmware, or a combination of hardware and software. As used herein, a processor is implemented in hardware, firmware, or a combination of hardware and software. As used herein, the phrase “based on” is intended to be broadly construed to mean “based at least in part on. ”
[0208] As used herein, a phrase referring to a list of items separated by “or” refers to any combination of those items, including single members. For example, “a, b, or c” is intended to cover the possibilities of: a only, b only, c only, a combination of a and b, a combination of a and c, a combination of b and c, and a combination of a and b and c.
[0209] In this disclosure, an expression of “X / Y” may include meaning of any of the following: “X or Y” or “X and Y” or “X and / or Y. " An expression of “ (A) B” or “B (A) ” may include concept of “only B. ” An expression of “ (A) B” or “B (A) ” may include the concept of “A+B” or “B+A. ”
[0210] In this disclosure, the term "can" indicates a capability, or alternatively indicates a possible implementation option. The term "may" indicates a permission or a possible implementation option.
[0211] Some aspects are described herein in connection with thresholds. As used herein, satisfying a threshold may refer to a value being greater than the threshold, greater than or equal to the threshold, less than the threshold, less than or equal to the threshold, equal to the threshold, not equal to the threshold, or the like.
[0212] The various illustrative components, logic, logical blocks, modules, circuits, operations and algorithm processes described in connection with the implementations disclosed herein may be implemented as electronic hardware, firmware, software, or combinations of hardware, firmware or software, including the structures disclosed in this specification and the structural equivalents thereof. The interchangeability of hardware, firmware and software has been described generally, in terms of functionality, and illustrated in the various illustrative components, blocks, modules, circuits and processes described above. Whether such functionality is implemented in hardware, firmware or software depends upon the particular application and design constraints imposed on the overall system.
[0213] As described above, some aspects of the subject matter described in this specification can be implemented as software. For example, various functions of components disclosed herein, or various blocks or steps of a method, operation, process or algorithm disclosed herein can be implemented as one or more modules of one or more computer programs. Such computer programs can include non-transitory processor-executable or computer-executable instructions encoded on one or more tangible processor-readable or computer-readable storage media for execution by, or to control the operation of, a data processing apparatus including the components of the devices described herein. By way of example, and not limitation, such storage media may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that may be used to store program code in the form of instructions or data structures. Combinations of the above should also be included within the scope of storage media.
[0214] Various modifications to the implementations described in this disclosure may be readily apparent to persons having ordinary skill in the art, and the generic principles defined herein may be applied to other implementations without departing from the scope of this disclosure. Thus, the claims are not intended to be limited to the implementations shown herein but are to be accorded the widest scope consistent with this disclosure, the principles and the novel features disclosed herein.
[0215] Additionally, various features that are described in this specification in the context of separate implementations also can be implemented in combination in a single implementation. Conversely, various features that are described in the context of a single implementation also can be implemented in multiple implementations separately or in any suitable subcombination. As such, although features may be described above as acting in particular combinations, and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.
[0216] The drawings may schematically depict one or more example processes in the form of a flowchart or flow diagram. However, other operations that are not depicted can be incorporated in the example processes that are schematically illustrated. For example, one or more additional operations can be performed before, after, simultaneously, or between any of the illustrated operations. In some circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the implementations described above should not be understood as requiring such separation in all implementations, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products. Additionally, other implementations are within the scope of the following claims. In some implementations, the actions recited in the claims can be performed in a different order and still achieve desirable results.
[0217] The foregoing disclosure provides illustration and description but is not intended to be exhaustive or to limit the aspects to the precise form disclosed. Modifications and variations may be made in light of the above disclosure or may be acquired from practice of the aspects. While the aspects of the disclosure have been described in terms of various examples, any combination of aspects from any of the examples is also within the scope of the disclosure. The examples in this disclosure are provided for pedagogical purposes.
Claims
A method for wireless communication by a user equipment (UE) (102) , comprising:receiving (330, 1230, 1430) , from a network entity (104) , a configuration (130, 630) that includes a plurality of channel state information (CSI) report configurations for a plurality of CSI reports, the plurality of CSI reports including one or more predictive CSI reports;generating (150) , by the UE, a predictive CSI report (161) of the one or more predictive CSI reports based on a priority for the predictive CSI report from among the plurality of CSI reports and a number of CSI processing units (CPUs) available for one or more prediction tasks associated with the predictive CSI report; andtransmitting (160, 360, 1260, 1460) , to the network entity, the predictive CSI report.The method of claim 1, further comprising calculating (351, 1251, 1451) the priority based on one or more of:whether the predictive CSI report includes a measured CSI value,a timestamp associated with a predicted CSI value for the predictive CSI report,a prediction function priority value associated with a prediction function, ora weighting value associated with the prediction function;wherein a machine learning (ML) model of a plurality of ML models maintained at the UE is used to generate the predicted CSI value for the predictive CSI report.The method of claim 1 or 2, further comprising:receiving (340, 1240) , from the network entity, one or more downlink reference signals (DL-RSs) ; andpredicting a CSI value for the predictive CSI report based on an output of a machine learning (ML) model, wherein an input to the ML model includes one or more of:measurements of the one or more DL-RSs;positional encoding associated with the one or more DL-RSs; ora task sequence associated with the one or more DL-RSs, wherein the ML model is trained to process a plurality of different task sequences.The method of any one of claims 1 to 3, wherein the UE maintains a single ML model for generating multiple predictive CSI reports.The method of any one of claims 1 to 4, further comprising reporting, to the network entity, the number of CPUs available for the generating the one or more predictive CSI reports.The method of any one of claims 1 to 5, further comprising reporting, to the network entity, a required number of CPUs for generating the one or more predictive CSI reports based on a computation capability of the UE.The method of claim 6, further comprisingupdating the required number of CPUs based on at least one of:a first change in power availability,a second change in a hardware configuration of the UE, ora third change in a hardware availability of the UE; andreporting, to the network entity, the updated required number of CPUs.The method of any one of claims 1 to 7, further comprising transmitting (320, 1220) , to the network entity, UE capability information that includes one or more of:one or more indicators indicating supported CSI prediction tasks,a first number of CPUs for the one or more CSI prediction tasks,a second number of CPUs for one or more prediction task types,a third number of CPUs for a type of inference configuration,a fourth number of CPUs for processing an ML model, ora fifth number of CPUs for a functionality of the ML model.The method of any one of claims 1 to 8, wherein the configuration includes one or more of:one or more ML functionalities,activated ML models, oractivated CSI reports.The method of any one of claims 1 to 9, further comprising:maintaining a plurality of ML models;selecting an ML model of the plurality of ML models based on at least one prediction task for a predicted value of the predictive CSI report; andincluding an indication of the selected ML model in the predictive CSI report.The method of any one of claims 1 to 10, further comprising:grouping the one or more predictive CSI reports into one or more CSI report groups; andtransmitting 324, 1224) , to the network entity, CSI grouping information that indicates the one or more CSI report groups.The method of claim 11, wherein the CSI grouping information includes a required number of CPUs corresponding to each CSI report group of the one or more CSI report groups.The method of claim 11 or 12, further comprising:transmitting, to the network entity, CSI group information that indicates at least one CSI report group of the one or more CSI report groups in use by the UE.The method of claim 13, wherein the CSI group information indicates a required number of CPUs for the at least one CSI report group in use by the UE.The method of any one of claims 11 to 14, wherein the priority comprises a first priority, the method further comprising:calculating a second priority for at least one CSI report group of the one or more CSI report groups;transmitting the one or more predictive CSI reports in the at least one CSI report group based on the second priority.The method of any one of claims 1 to 15, wherein the priority of the predictive CSI report comprises a first priority, the method further comprising:refraining from generating a second predictive CSI report of the plurality of CSI reports when a second priority for the at least one CSI report is lower than the first priority and the number of required CPUs corresponding to the second predictive CSI report exceeds the number of available CPUs.The method of any one of claims 1 to 16, wherein the priority comprises a first priority, the method further comprising:calculating a second priority for at least one other CSI report of the plurality of CSI reports, andtransmitting, to the network entity, the at least one other CSI report based on the second priority.A method for wireless communication by a network entity (104) , comprising:transmitting (330, 1330, 1530) , to a user equipment (UE) (102) , a configuration (130, 630) that includes:a plurality of channel state information (CSI) report configurations for a plurality of CSI reports, the plurality of CSI reports including one or more predictive CSI reports, anda prediction function priority for a prediction function used to predict a CSI value for the one or more predictive CSI reports;transmitting (340, 1340, 1540) , to the UE, a plurality of downlink reference signals; andreceiving (160, 360, 1360, 1560) , from the UE, at least one predictive CSI report including the predicted CSI value based on the plurality of downlink reference signals.The method of claim 18, further comprising:receiving, from the UE, UE capability information including one or more of:one or more indicators indicating supported CSI prediction tasks,a first number of CSI processing units (CPUs) for the one or more CSI prediction tasks,a second number of CPUs for one or more prediction task types,a third number of CPUs for a type of inference configuration,a fourth number of CPUs for processing an ML model, ora fifth number of CPUs for a functionality of the ML model; andselecting a predictive CSI report to include in the configuration based on the UE capability information.An apparatus, comprising:a communication unit; anda processing system configured to control the communication unit to implement any one of the methods of any one of claims 1 to 19.