Communication method and device

By executing model quantization configuration information in communication devices, the AI ​​model has solved the problem of high computing density and high power consumption in terminal scenarios, and the model size is reduced and the inference efficiency is improved.

WO2024234267A1PCT designated stage expired Publication Date: 2024-11-21GUANGDONG OPPO MOBILE TELECOMMUNICATIONS CORP LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2023/094366
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-15
Publication Date
2024-11-21

AI Technical Summary

Technical Problem

In the field of communication, AI models need to be as small as possible in terminal scenarios and have high throughput. The existing technology has failed to effectively quantify the model, resulting in high computing density, large power consumption and low inference efficiency.

Method used

By quantizing the configuration information of the model, the communication device is instructed to perform model quantization-related operations, such as converting the parameter accuracy of the AI ​​model from 32-bit floating point to low precision, such as INT8, reducing the model size and accelerating the calculation.

Benefits of technology

The model size is reduced, hardware memory and storage usage is reduced, device power consumption is reduced, model inference efficiency is improved, and air interface transmission overhead is reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2023094366_21112024_PF_FP_ABST
    Figure CN2023094366_21112024_PF_FP_ABST
Patent Text Reader

Abstract

The present application relates to a communication method and device. The method comprises: a first communication device receives model quantization configuration information from a second communication device; and the first communication device executes a model quantization related operation according to the model quantization configuration information. In embodiments of the present application, the model quantization configuration information can indicate a model quantization related operation, and model quantization is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Communication method and device Technical Field

[0001] The present application relates to the field of communications, and more specifically, to a communication method and device. Background Art

[0002] The use of artificial intelligence (AI) technology has already begun in the communications field. The application of AI models in communications fields, such as wireless communications, requires consideration of actual scenario requirements. For example, to be suitable for terminal scenarios, the model may require to be as small as possible and have a high throughput. Therefore, AI model quantization is necessary.

[0003] Summary of the Invention

[0004] The embodiments of the present application provide a communication method and device that can implement model quantization.

[0005] An embodiment of the present application provides a communication method, including:

[0006] The first communication device receives model quantization configuration information from the second communication device;

[0007] The first communication device performs model quantization related operations according to the model quantization configuration information.

[0008] An embodiment of the present application provides a communication method, including:

[0009] The second communication device sends model quantization configuration information to the first communication device, where the model quantization configuration information is used to instruct the second communication device to perform model quantization related operations.

[0010] An embodiment of the present application provides a first communication device, including:

[0011] A receiving unit, configured to receive model quantization configuration information from a second communication device;

[0012] A processing unit is used to perform model quantization related operations according to the model quantization configuration information.

[0013] An embodiment of the present application provides a second communication device, including:

[0014] A sending unit is used to send model quantization configuration information to the first communication device, where the model quantization configuration information is used to instruct the second communication device to perform model quantization related operations.

[0015] An embodiment of the present application provides a communication device, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to call and execute the computer program stored in the memory, so that the communication device executes the above-mentioned communication method.

[0016] The embodiment of the present application provides a chip for implementing the above-mentioned communication method. Specifically, the chip includes: a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the above-mentioned communication method.

[0017] An embodiment of the present application provides a computer-readable storage medium for storing a computer program, which, when executed by a device, enables the device to execute the above-mentioned communication method.

[0018] An embodiment of the present application provides a computer program product, including computer program instructions, which enable a computer to execute the above-mentioned communication method.

[0019] An embodiment of the present application provides a computer program, which, when executed on a computer, enables the computer to execute the above-mentioned communication method.

[0020] In the embodiment of the present application, model quantization configuration information can be used to indicate related operations of model quantization to achieve model quantization. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] FIG1 is a schematic diagram of an application scenario according to an embodiment of the present application.

[0022] Figure 2 is a schematic diagram of a multi-layer neural network.

[0023] FIG3 is a schematic diagram of a channel state information feedback system.

[0024] 4a and 4b are schematic diagrams of the beam scanning process.

[0025] FIG5 is a schematic flowchart of a communication method according to an embodiment of the present application.

[0026] FIG6 is a schematic flowchart of a communication method according to another embodiment of the present application.

[0027] FIG7 is a schematic flowchart of a communication method according to another embodiment of the present application.

[0028] FIG8 is a schematic flowchart of a communication method according to an embodiment of the present application.

[0029] FIG9 is a schematic flowchart of a communication method according to another embodiment of the present application.

[0030] Figure 10 is a schematic flowchart of an AI model quantization indication method according to an embodiment of the present application.

[0031] FIG11 is a schematic flowchart according to Example 1-1 of the present application.

[0032] FIG12 is a schematic flow chart according to Examples 1-2 of the present application.

[0033] FIG13 is a schematic flowchart according to Example 2-1 of the present application.

[0034] FIG14 is a schematic flowchart according to Example 2-2 of the present application.

[0035] FIG15 is a schematic flow chart according to Example 3 of the present application.

[0036] FIG16 is a schematic block diagram of a first communication device according to an embodiment of the present application.

[0037] FIG17 is a schematic block diagram of a first communication device according to another embodiment of the present application.

[0038] FIG18 is a schematic block diagram of a second communication device according to an embodiment of the present application.

[0039] FIG19 is a schematic block diagram of a second communication device according to another embodiment of the present application.

[0040] Figure 20 is a schematic block diagram of a communication device according to an embodiment of the present application.

[0041] FIG21 is a schematic block diagram of a chip according to an embodiment of the present application.

[0042] Figure 22 is a schematic block diagram of a communication system according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0044] The technical solutions of the embodiments of the present application can be applied to various communication systems, such as: Global System of Mobile communication (GSM) system, Code Division Multiple Access (CDMA) system, Wideband Code Division Multiple Access (WCDMA) system, General Packet Radio Service (GPRS), Long Term Evolution (LTE) system, Advanced Long Term Evolution (LTE-A) system, New Radio (NR) system, NR system evolution system, LTE on unlicensed spectrum (LTE-U) system, NR on unlicensed spectrum (NR-U) system, Non-Terrestrial Networks (NTN) system, Universal Mobile Telecommunication System (UMTS), Wireless Local Area Networks (WLAN), Wireless Fidelity (Wireless Fidelity) system. Fidelity, WiFi), fifth-generation communication (5th-Generation, 5G) system or other communication systems, etc.

[0045] Generally speaking, traditional communication systems support a limited number of connections and are easy to implement. However, with the development of communication technology, mobile communication systems will not only support traditional communications, but will also support, for example, device-to-device (D2D) communication, machine-to-machine (M2M) communication, machine-type communication (MTC), vehicle-to-vehicle (V2V) communication, or vehicle-to-everything (V2X) communication, etc. The embodiments of the present application can also be applied to these communication systems.

[0046] In one embodiment, the communication system in the embodiment of the present application can be applied to a carrier aggregation (CA) scenario, a dual connectivity (DC) scenario, and a standalone (SA) networking scenario.

[0047] In one embodiment, the communication system in the embodiment of the present application can be applied to an unlicensed spectrum, wherein the unlicensed spectrum can also be considered as a shared spectrum; or, the communication system in the embodiment of the present application can also be applied to an authorized spectrum, wherein the authorized spectrum can also be considered as an unshared spectrum.

[0048] The embodiments of the present application describe various embodiments in conjunction with network devices and terminal devices, wherein the terminal device may also be referred to as user equipment (UE), access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent or user device, etc.

[0049] The terminal device can be a station (STAION, ST) in a WLAN, a cellular phone, a cordless phone, a Session Initiation Protocol (SIP) phone, a Wireless Local Loop (WLL) station, a Personal Digital Assistant (PDA) device, a handheld device with wireless communication capabilities, a computing device or other processing device connected to a wireless modem, a vehicle-mounted device, a wearable device, a terminal device in a next-generation communication system such as an NR network, or a terminal device in a future evolved Public Land Mobile Network (PLMN) network, etc.

[0050] In an embodiment of the present application, the terminal device can be deployed on land, including indoors or outdoors, handheld, wearable or vehicle-mounted; it can also be deployed on the water surface (such as ships, etc.); it can also be deployed in the air (such as airplanes, balloons and satellites, etc.).

[0051] In an embodiment of the present application, the terminal device may be a mobile phone, a tablet computer, a computer with wireless transceiver function, a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal device in industrial control, a wireless terminal device in self-driving, a wireless terminal device in remote medical, a wireless terminal device in a smart grid, a wireless terminal device in transportation safety, a wireless terminal device in a smart city, or a wireless terminal device in a smart home, etc.

[0052] As an example and not a limitation, in the embodiment of the present application, the terminal device may also be a wearable device. Wearable devices may also be called wearable smart devices, which are a general term for wearable devices that are intelligently designed and developed using wearable technology for daily wear, such as glasses, gloves, watches, clothing, and shoes. A wearable device is a portable device that is worn directly on the body or integrated into the user's clothes or accessories. Wearable devices are not only hardware devices, but also achieve powerful functions through software support, data interaction, and cloud interaction. Broadly speaking, wearable smart devices include those that are fully functional, large in size, and can achieve complete or partial functions without relying on smartphones, such as smart watches or smart glasses, as well as those that only focus on a certain type of application function and need to be used in conjunction with other devices such as smartphones, such as various smart bracelets and smart jewelry for vital sign monitoring.

[0053] In an embodiment of the present application, the network device may be a device for communicating with a mobile device. The network device may be an access point (AP) in WLAN, a base station (BTS) in GSM or CDMA, a base station (NodeB, NB) in WCDMA, an evolved base station (eNB or eNodeB) in LTE, or a relay station or access point, or a vehicle-mounted device, a wearable device, and a network device (gNB) in an NR network, or a network device in a future evolved PLMN network or a network device in an NTN network, etc.

[0054] As an example and not a limitation, in an embodiment of the present application, the network device may have a mobile feature, for example, the network device may be a mobile device. Alternatively, the network device may be a satellite or a balloon station. For example, the satellite may be a low earth orbit (LEO) satellite, a medium earth orbit (MEO) satellite, a geostationary earth orbit (GEO) satellite, a high elliptical orbit (HEO) satellite, etc. Optionally, the network device may also be a base station set up in a location such as land or water.

[0055] In an embodiment of the present application, the network device can provide services for a cell, and the terminal device communicates with the network device through the transmission resources used by the cell (for example, frequency domain resources, or spectrum resources). The cell can be a cell corresponding to the network device (for example, a base station). The cell can belong to a macro base station or a base station corresponding to a small cell. The small cells here may include: metro cells, micro cells, pico cells, femto cells, etc. These small cells have the characteristics of small coverage and low transmission power, and are suitable for providing high-speed data transmission services.

[0056] FIG1 exemplarily illustrates a communication system 100. The communication system includes a network device 110 and two terminal devices 120. In one embodiment, the communication system 100 may include multiple network devices 110, and each network device 110 may include a different number of terminal devices 120 within its coverage area, which is not limited in this embodiment of the present application.

[0057] In one embodiment, the communication system 100 may further include other network entities such as a Mobility Management Entity (MME) and an Access and Mobility Management Function (AMF), which is not limited in this embodiment of the present application.

[0058] Among them, the network equipment may include access network equipment and core network equipment. That is, the wireless communication system also includes multiple core networks for communicating with the access network equipment. The access network equipment can be an evolutionary base station (evolutional node B, abbreviated as eNB or e-NodeB) macro base station, micro base station (also called "small base station"), pico base station, access point (AP), transmission point (TP) or new generation base station (new generation Node B, gNodeB), etc. in a long-term evolution (LTE) system, a next-generation (mobile communication system) (next radio, NR) system or an authorized auxiliary access long-term evolution (LAA-LTE) system.

[0059] It should be understood that in the embodiments of the present application, a device having a communication function in a network / system may be referred to as a communication device. Taking the communication system shown in Figure 1 as an example, the communication device may include a network device and a terminal device having a communication function. The network device and the terminal device may be specific devices in the embodiments of the present application and will not be described in detail here. The communication device may also include other devices in the communication system, such as a network controller, a mobility management entity, and other network entities, which are not limited in the embodiments of the present application.

[0060] It should be understood that the terms "system" and "network" are often used interchangeably herein. The term "and / or" is simply a description of an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " generally indicates that the related objects are in an "or" relationship.

[0061] It should be understood that the "indication" mentioned in the embodiments of this application can be a direct indication, an indirect indication, or an indication of an association. For example, "A indicates B" can mean that A directly indicates B, for example, B can be obtained through A; it can also mean that A indirectly indicates B, for example, A indicates C, and B can be obtained through C; it can also mean that there is an association between A and B.

[0062] In the description of the embodiments of the present application, the term "corresponding" may indicate a direct or indirect correspondence between the two, or an association relationship between the two, or a relationship between indication and being indicated, configuration and being configured, etc.

[0063] To facilitate understanding of the technical solutions of the embodiments of the present application, the relevant technologies of the embodiments of the present application are described below. The following relevant technologies can be arbitrarily combined with the technical solutions of the embodiments of the present application as optional solutions, and they all fall within the protection scope of the embodiments of the present application.

[0064] 1. Artificial Intelligence

[0065] In recent years, artificial intelligence (AI) research, exemplified by neural networks, has achieved remarkable success in numerous fields and will continue to play a vital role in our lives and production for a long time to come. In particular, machine learning (ML), a key research area in AI, leverages the nonlinear processing capabilities of neural networks (NNs) to successfully solve a range of previously intractable problems. ML has recently garnered increasing attention, demonstrating superior performance to humans in areas such as image recognition, speech processing, natural language processing, and gaming.

[0066] One type of training model in AI is an artificial neural network. AI networks create models based on training data, typically to generate more accurate predictions. Neural networks typically use a multi-layer design. As shown in Figure 2, layer-by-layer training of multi-layer neural networks for feature learning significantly improves the learning and processing capabilities of neural networks. This has led to widespread applications in pattern recognition, signal processing, optimization and combination, and anomaly detection.

[0067] AI has important application potential in many aspects, including complex and unknown environment modeling, learning, channel prediction, intelligent signal generation and processing, network status tracking and intelligent scheduling, and network optimization and deployment. It is expected to promote the evolution of future communication paradigms and changes in network architecture, and is of great significance and value to 6G technology research.

[0068] 2. AI model quantification

[0069] In terminal scenarios, memory and speed are generally considered, so the model must be as small as possible while maintaining high throughput. Quantization can be used to reduce model size and speed up the model.

[0070] The quantization process primarily converts trained deep neural network weights and activation function outputs from high-precision to low-precision. For example, a model trained in floating-point FP32 format can be compressed to fixed-point INT8 (or INT4 / INT1). Since INT8 only requires 8 bits for representation, the model size can theoretically be reduced to 1 / 4 of the original size compared to 32-bit floating-point. This compression ratio is very intuitive. Furthermore, most terminal devices have dedicated fixed-point computing units, and low-precision operators implemented using low-bit instructions can significantly improve speed. Of course, this also relies on collaborative architectures and algorithms to achieve even greater acceleration.

[0071] Quantization techniques can be primarily categorized into quantization-aware training (QAT) and post-training quantization (PTQ). The primary difference between the two is that quantization training involves inserting pseudo-quantization operators (which simulate the logic of low-precision operations) into the model and fine-tuning the original floating-point model through optimization methods such as gradient descent. This fine-tuning ultimately results in a model with the desired accuracy. Offline quantization primarily uses a small calibration dataset to determine the network distribution. Quantization parameters are then obtained through statistical methods or optimization of the distributions of floating-point and fixed-point outputs, ultimately yielding the final deployed model. Both approaches have their advantages and disadvantages. Quantization training relies on the training logic of the original floating-point model, theoretically ensuring convergence to the accuracy of the original model. However, this requires careful parameter tuning and a longer production cycle. Offline quantization, however, relies on only a small amount of calibration data, resulting in a shorter production cycle and greater flexibility. However, its accuracy may be slightly lower than that of quantization training. During the actual implementation process, it was found that most models can achieve good model accuracy through offline quantization (accuracy loss within 1%, of course, this part of the accuracy improvement is also due to the support of the optimization strategy). The remaining few models may need to be trained through quantization to compensate for the accuracy loss. Therefore, in actual business, the advantages and disadvantages of both are combined for application.

[0072] 3. AI Channel State Information (CSI) Feedback in NR Systems

[0073] Given the tremendous success of AI technology, especially deep learning, in computer vision and natural language processing, the communications field has begun to explore the use of deep learning to solve technical challenges that are difficult to address with traditional communications methods. For example, the neural network architecture commonly used in deep learning is nonlinear and data-driven. It can extract features from the actual channel matrix data and restore the channel matrix information compressed and fed back by the UE as much as possible on the base station side. This not only ensures the restoration of channel information, but also provides the possibility of reducing the CSI feedback overhead on the UE side. Deep learning-based CSI feedback treats channel information as an image to be compressed, uses a deep learning autoencoder to compress and feed back the channel information, and reconstructs the compressed channel image at the transmitter, which can preserve the channel information to a greater extent.

[0074] Figure 3 shows a channel state information feedback system. This system is an autoencoder architecture consisting of an encoder and a decoder, deployed at the transmitter and receiver, respectively. After obtaining channel information through channel estimation, the transmitter compresses and encodes the channel information matrix using the encoder's neural network and feeds the compressed bitstream back to the receiver via an air interface feedback link. The receiver uses the decoder to recover the channel information from the feedback bitstream, obtaining complete feedback channel information. The network model structures within the encoder and decoder shown in Figure 3 allow for flexible design.

[0075] 4. AI Beam Management in NR Systems

[0076] The first version of the NR system, Rel.15, introduced millimeter wave frequency band communications, that is, the corresponding beam management mechanism was introduced. To put it simply, it can be divided into uplink and downlink beam management. Because the solution of the embodiment of the present application mainly focuses on the quantization operation in the model transmission, there is no essential difference between uplink and downlink for model transmission, so the downlink beam management mechanism is mainly introduced. The mechanism includes downlink beam scanning, UE optimal beam reporting, network (NewWork, NW) indication of downlink beams and other processes.

[0077] The downlink beam scanning process may include: the NW scans different transmit beam directions using the downlink reference signal synchronization signal and the Physical Broadcast Channel (PBCH) block (SSB) and / or the Channel State Information-Reference Signal (CSI-RS). The UE can use different receive beams for measurement, thereby traversing all beam pair combinations, and the UE calculates its L1 (Layer 1) Reference Signal Received Power (RSRP) value.

[0078] Figures 4a and 4b are schematic diagrams of the beam scanning process. Figure 4a shows traversing the transmit beam and receive beam, while Figure 4b shows traversing the receive beam for a specific transmit beam.

[0079] Beam reporting may include: the UE compares the L1-RSRP values ​​of all measured beam pairs, selects the K transmit beams with the highest L1-RSRP, and then reports them to the NW as uplink control information. Here, L1-RSRP can also be replaced by other beam link indicators, such as L1-Signal to Interference plus Noise Ratio (SINR), L1-Reference Signal Receiving Quality (RSRQ), etc. After decoding the UE's beam reporting, the NW can carry the Transmission Configuration Indicator (TCI) status (including the transmit beam with SSB or CSI-RS as a reference) through Media Access Control (MAC) or Downlink Control Information (DCI) signaling to complete the beam indication to the UE. The UE uses the receive beam corresponding to the transmit beam for reception.

[0080] AI-based beam management is one of the main use cases of AI projects. An implementation framework for AI-based spatial and temporal beam prediction includes:

[0081] Beam prediction is performed on beam set A using the measurement results of beam set B.

[0082] Set B may be a subset of Set A, or Set B and Set A may be different beam sets (for example, Set A uses a narrow beam, and Set B uses a wide beam);

[0083] AI models can be deployed on the NW side or the UE side;

[0084] The measurement result on set B can be L1-RSRP, or other auxiliary information, such as beam (pair) ID, etc.

[0085] Although R18 discussed and studied technical solutions for AI model enhancement, including different AI model training processes and AI model transmission methods, there was no detailed discussion on AI model quantization. The parameter accuracy of the AI ​​model depends on the training framework. Generally, the default precision type of the model parameters saved by the framework is FLOPP32 (32-bit single-precision floating-point type). AI models with high-precision parameters mean higher memory usage and lower computing density, and the inference process requires more memory accesses. The bottleneck of AI model inference efficiency is generally not computing but memory access. In this case, if AI model quantization is not considered, lower computing density will be relatively more time-consuming. At the same time, power consumption mainly comes from computing and memory access. The computing energy consumption before and after model quantization is an order of magnitude different.

[0086] In summary, before AI solutions are widely implemented in the wireless field, they must fully consider the actual scenario requirements and ensure that the loss of inference accuracy is within the required range and the model is as small as possible. Therefore, effective model quantization is necessary to address the challenges of limited system hardware resources.

[0087] FIG5 is a schematic flow chart of a communication method according to an embodiment of the present application. The method 500 can optionally be applied to the system shown in FIG1 , but is not limited thereto. The method includes at least part of the following contents.

[0088] S510. The first communication device receives model quantization configuration information from the second communication device;

[0089] S520: The first communication device performs model quantization-related operations according to the model quantization configuration information.

[0090] In an embodiment of the present application, the model may be an AI model. For example, a CSI prediction model, a beam prediction model, a UE positioning model, a channel estimation model, and the like. The first communication device may be a terminal device or a network device. The second communication device may be a terminal device or a network device. If the first communication device and the second communication device have an associated or identical AI model, the second communication device may send model quantization configuration information to the first communication device. The model quantization configuration information may include information related to the quantization of the AI ​​model in the first communication device. For example, one or more of the model's name, function, parameters to be quantized, quantization accuracy information, and the like.

[0091] In the embodiments of the present application, model quantization configuration information can be used to indicate model quantization-related operations, effectively implementing model quantization. For example, quantized models or model parameters can be transmitted between communication devices through model quantization configuration information. The size of the quantized AI model can be reduced exponentially, significantly reducing air interface transmission overhead, hardware memory and storage usage, device power consumption, and improving model inference efficiency.

[0092] Figure 6 is a schematic flow chart of a communication method according to another embodiment of the present application. Method 600 may include one or more features of the communication method of the above embodiment. In one embodiment, the model quantization configuration information includes target accuracy information. The target accuracy information in the model quantization configuration information sent by the second communication device can be understood as the model accuracy requirement of the second communication device.

[0093] In an embodiment of the present application, the initial parameter accuracy of the AI ​​model depends on the training framework. For example, the training accuracy type of some models can be a 32-bit single-precision floating-point type or other types. The parameter accuracy that can be obtained by quantizing the trained AI model may be different from the initial parameter accuracy, for example, lower than the initial parameter accuracy. The same AI model uses different quantization methods and / or quantization accuracy, and can be converted into AI models with different parameter accuracy. For example, the quantization accuracy types of the quantized AI model include signed 8-bit integer (Int8) and unsigned 8-bit integer (UInt8).

[0094] In an embodiment of the present application, the target accuracy information included in the model quantization configuration information can be used to find out whether there is a model that meets the requirements in the first communication device. The target accuracy information may include a target accuracy type such as Int8 or UInt8, and may also include indication information corresponding to the target accuracy type. For example, 0 indicates Int8, 1 indicates UInt8; or 1 indicates Int8, 0 indicates UInt8. A target accuracy type may correspond to one or more parameters of a model.

[0095] In one embodiment, the target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

[0096] In embodiments of the present application, an AI model may include multiple types of parameters, and the types of parameters that can be quantified in different AI models may be the same or different. For example, the parameters of an AI model may include, but are not limited to, one or more of an activation parameter, a bias parameter, and a weight parameter. The AI ​​model may be quantized based on one or more of the activation parameters, bias parameters, and weight parameters to obtain one or more quantized models.

[0097] If the model quantization configuration information received by the first communication device includes target accuracy information, it can be checked whether there is an AI model corresponding to the target accuracy information. There are many ways to indicate the target accuracy information, as shown below.

[0098] In method 1, model parameters do not need to be divided into weight parameters and activation parameters. The following quantization data types are supported, and 1 bit can be used in the AI ​​Model Quantization Indicator Configuration (AIMQIC) to indicate the target accuracy information. The following example shows:

[0099] 0: Int8 (QuantType.QInt8): signed 8-bit integer

[0100] 1: UInt8 (QuantType.QUInt8): unsigned 8-bit integer

[0101] In this method, the types corresponding to 0 and 1 are interchangeable. There is no need to divide the model into multiple parts, and all parameters can support the same quantization accuracy.

[0102] In the second method, the model parameters are divided into weight parameters (activation), bias parameters (bias), and activation parameters (weight). The following quantization data types are supported, and AIMQIC can use 3 bits to indicate the target accuracy information:

[0103] Int8 (QuantType.QInt8): signed 8-bit integer

[0104] UInt8 (QuantType.QUInt8): unsigned 8-bit integer.

[0105] Among them, AIMQIC uses 3 bits to indicate target accuracy information, as shown below:

[0106] 000: U8U8U8, (activation: uint8, bias: uint8, weight: uint8)

[0107] 001: U8U8S8, (activation: uint8, bias: uint8, weight: int8)

[0108] 010: U8S8U8, (activation: uint8, bias: int8, weight: uint8)

[0109] 011: U8S8S8, (activation: uint8, bias: int8, weight: int8)

[0110] 100: S8U8U8, (activation: int8, bias: uint8, weight: uint8)

[0111] 101: S8U8S8, (activation: int8, bias: uint8, weight: int8)

[0112] 110: S8S8U8, (activation: int8, bias: int8, weight: uint8)

[0113] 111: S8S8S8, (activation: int8, bias: int8, weight: int8)

[0114] In this way, the 3-bit value of AIMQIC and its corresponding parameter precision type can be changed, dividing the model into several detailed parts, each supporting different quantization precisions.

[0115] Method 3: Model parameters are not divided into one or more of weight parameters (activation), bias parameters (bias), and activation parameters (weight), and supports the following multiple quantization data types. In this method, AIMQIC can use 2 bits to indicate target accuracy information, as shown in the following example:

[0116] 00: Int8 (QuantType.QInt8): signed 8-bit integer

[0117] 01: UInt8 (QuantType.QUInt8): unsigned 8-bit integer

[0118] 10: Float8 (QuantType.QFloat8): 8-bit single-precision floating point type

[0119] 11: Float16 (QuantType.QFloat16): 16-bit single-precision floating point type

[0120] In this way, the 2-bit value of AIMQIC and its corresponding parameter precision type can be changed, and fewer or more bits can represent coarser or more detailed model quantization accuracy requirements.

[0121] In addition, AIMQIC can indicate the target accuracy information of weight parameters, bias parameters, or activation parameters separately. AIMQIC can also indicate the target accuracy information of any two of weight parameters, bias parameters, and activation parameters.

[0122] In one embodiment, the model quantization accuracy requirement supported by the first communication device may include global quantization accuracy.

[0123] In one embodiment, the model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy, and weight parameter quantization accuracy.

[0124] In an embodiment of the present application, the model quantization accuracy may include quantization accuracy related to one or more model parameters, or may be the global quantization accuracy of the model.

[0125] For example, model A is quantized to obtain model A1 and model A2, where the quantization accuracy of model A1 is Int8 and the quantization accuracy of model A2 is UInt8. Model A, model A1, and model A2 can be models with the same function but different parameters.

[0126] For example, model B is quantized to obtain models B1, B2, and B3. The activation parameters of model B1 are quantized to Int8 precision, and the weight parameters are quantized to Float8 precision. The activation parameters of model B2 are quantized to UInt8 precision, and the weight parameters are quantized to Float8 precision. The activation parameters of model B3 are quantized to UInt8 precision, and the weight parameters are quantized to Float16 precision. Models B, B1, B2, and B3 can have the same functionality but different parameters.

[0127] In one embodiment, as shown in FIG6 , step S520 further includes:

[0128] S610. When the first communication device includes a first model corresponding to the target accuracy information, send the first model and / or parameters of the first model to the second communication device.

[0129] For example, the model quantization configuration information received by the first communication device includes model name information A and target accuracy information Int8 corresponding to indicator bit 1. The first communication device finds model A1 (quantization accuracy Int8) that meets the target accuracy information in the model quantization configuration information.

[0130] For example, the model quantization configuration information received by the first communication device includes model name information B, quantization parameters including activation parameters and weight parameters, target quantization precision of the activation parameters being Int8, and target quantization precision of the weight parameters being Float8. Model B2 (the quantization precision of the activation parameters being UInt8, and the quantization precision of the weight parameters being Float8) that meets the target precision information in the model quantization configuration information is found in the first communication device.

[0131] In one embodiment, as shown in FIG6 , step S520 further includes:

[0132] S620: If the first communication device does not include a first model corresponding to the target accuracy information, convert the second model into the first model according to the target accuracy information;

[0133] S630: The first communication device sends the first model and / or parameters of the first model to the second communication device.

[0134] If the model quantization accuracy requirements supported by the first communication device do not include requirements that comply with the target accuracy information, the first communication device can first quantize the model based on the target accuracy information to obtain a model that complies with the target accuracy information. Then, the model that complies with the target accuracy information is sent to the second communication device. For example, the model quantization configuration information includes model name information A and target quantization accuracy Float8. However, the first communication device only has model A1 with a quantization accuracy of Int8 and model A2 with a quantization accuracy of UInt8. In this case, model A can be converted to model A3 with a quantization accuracy of Float8.

[0135] After obtaining the first model corresponding to the target accuracy information, the first communication device can send the first model corresponding to the target accuracy information, or the parameters of the first model, to the second communication device. For example, if the second communication device does not have a model for a certain function, the first communication device can send the first model of the function to the second communication device. If the second communication device has a model for the function, the first communication device can send the parameters of the first model to the second communication device to update the parameters of the original model.

[0136] FIG7 is a schematic flow chart of a communication method according to another embodiment of the present application. The method 700 may include one or more features of the communication method of the above embodiment. In one implementation, as shown in FIG7 , the method further includes:

[0137] S710. The first communication device sends a model quantization configuration request to the second communication device.

[0138] In an embodiment of the present application, the first communication device may first send a model quantization configuration request to the second communication device, and the second communication device may return model quantization configuration information in response to the model quantization configuration request. The second communication device may also actively send model quantization configuration information, see S510.

[0139] In one embodiment, the first model is obtained based on a model trained by the first communication device or a sub-model of a model trained by the first communication device. In an embodiment of the present application, the first model sent by the first communication device to the second communication device can be a quantized model based on the trained complete model, or a quantized model based on a sub-model of the trained complete model. For example, the trained complete model is a CSI model, and the CSC model includes an encoder model and a decoder model.

[0140] In one implementation, the first communication device is a network device, and the second communication device is a terminal device.

[0141] In one embodiment, the first model is an encoder model of a CSI model, and the CSI model is trained on the network device and deployed on the network device and / or the terminal device. For example, the encoder model and decoder model of the CSI model are trained on the network device and then quantized. If the target accuracy information of the CSI model received by the network device has a corresponding encoder model, the encoder model or its parameters may be sent to the terminal device.

[0142] In one embodiment, the first communication device is a terminal device, and the second communication device is a network device.

[0143] In one embodiment, the first model is a decoder model of a CSI model, and the CSI model is trained on the terminal device and deployed on the network device and / or the terminal device. For example, the encoder model and decoder model of the CSI model are trained on the terminal device and then quantized. If the target accuracy information of the CSI model received by the terminal device has a corresponding decoder model, the decoder model or its parameters may be sent to the terminal device.

[0144] In one embodiment, the first communication device is a first terminal device, and the second communication device is a second terminal device.

[0145] In one embodiment, the first model is an encoder model and / or decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device. For example, the encoder model and decoder model of the CSI model are trained on the first terminal device and the model is quantized. If the target accuracy information of the CSI model received by the first terminal device has a corresponding encoder model and / or decoder model, the encoder model and / or decoder model can be sent to the second terminal device, and the parameters of the encoder model and / or decoder model can also be sent to the second terminal device.

[0146] FIG8 is a schematic flow chart of a communication method according to an embodiment of the present application. The method 800 can optionally be applied to the system shown in FIG1 , but is not limited thereto. The method includes at least part of the following contents.

[0147] S810. The second communication device sends model quantization configuration information to the first communication device, where the model quantization configuration information is used to instruct the second communication device to perform model quantization-related operations.

[0148] Fig. 9 is a schematic flow chart of a communication method according to another embodiment of the present application. The method 900 may include one or more features of the communication method of the above embodiment. In one embodiment, the model quantization configuration information includes target accuracy information.

[0149] In one embodiment, the target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

[0150] In one embodiment, as shown in FIG9 , the method further includes:

[0151] S910. The second communication device receives a model quantization configuration request sent by the first communication device.

[0152] In one embodiment, as shown in FIG9 , the method further includes:

[0153] S920. The second communication device receives a first model corresponding to the target accuracy information sent by the first communication device.

[0154] In one embodiment, the model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy, and weight parameter quantization accuracy.

[0155] In one embodiment, the first model is obtained based on a model trained by the first communication device or a sub-model of the model trained by the first communication device.

[0156] In one implementation, the first communication device is a network device, and the second communication device is a terminal device.

[0157] In one embodiment, the first model is an encoder model of a CSI model, the CSI model is trained on the network device, and is deployed on the network device and / or the terminal device.

[0158] In one embodiment, the first communication device is a terminal device, and the second communication device is a network device.

[0159] In one embodiment, the first model is a decoder model of a CSI model, the CSI model is trained on the terminal device, and is deployed on the network device and / or the terminal device.

[0160] In one embodiment, the first communication device is a first terminal device, and the second communication device is a second terminal device.

[0161] In one embodiment, the first model is an encoder model and / or a decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device.

[0162] For a specific example of the second communication device executing the communication method 900 in this embodiment, reference can be made to the relevant description of the second communication device in the embodiment of the communication method executed by the first communication device described above, which will not be repeated here for the sake of brevity.

[0163] The communication method of an embodiment of the present application may include a method for indicating AI model quantization, as shown in Figure 10. During the AI ​​model transmission process, before the transmitting end sends the AI ​​model to the receiving end, the receiving end may send configuration information to the transmitting end to implement the indication of model quantization. For example, the AI ​​model quantization indication configuration (AIMQIC) reporting process in the AI ​​framework workflow can be performed before the AI ​​model is deployed. In the method for indicating AI model quantization, one or more of the following steps may be included.

[0164] S1010: The transmitting entity first reports an AI model quantization configuration request to the receiving entity (optional step). The model quantization configuration may also be referred to as a model quantization indication.

[0165] S1020, the receiving entity reports AI model quantization configuration information (AI Model Quantization Indicator Configuration, AIMQIC) to the sending entity.

[0166] At step S1030, the originating entity performs model quantization-related operations based on the received model quantization configuration information, querying an AI model of the target precision type. If an AI model of the target precision type exists, the process proceeds directly to step S1040. If no AI model of the target precision type is found, the process performs AI model quantization, converting the existing AI model of the target precision type into a quantized AI model.

[0167] S1040: Send the AI ​​model of the target accuracy type to the receiving entity.

[0168] At S1050, the receiving end receives the AI ​​model that meets the AIMQIC requirements in S1020 and completes the deployment and subsequent reasoning tasks.

[0169] The above process can be applied to different use cases, and the specific reporting content can also take various forms. The following is a detailed introduction to the different AIMQIC forms under different model transmission schemes.

[0170] Example 1: AI model quantization in a one-sided training and two-sided deployment AI solution

[0171] Example: 1-1: NW-side training, with UE and NW deployed separately. This example uses AI channel state information feedback as an example.

[0172] In one example, the basic network structure of AI channel state information feedback may include a model structure of an encoder and a decoder.

[0173] This example proposes an AI model quantization indication method in a channel state feedback scheme trained only on the NW (Network) side, as shown in Figure 11.

[0174] In this example, the CSI models (encoder and decoder models) are all trained on the NW side. The encoder model needs to be deployed on the UE side, so the NW needs to send the encoder model to the UE. Before sending the encoder model, it is necessary to report the AI ​​Model Quantization Indication Configuration (AIMQIC) to request a quantized encoder model. The specific process example is as follows:

[0175] At step S1110, the NW sends an AI model quantization configuration request to the UE via DCI. This step is optional and can also be triggered by the CSI service to directly trigger the UE to execute step S1120.

[0176] S1120: The UE reports AI model quantization configuration information (AIMQIC) to the NW via UCI. Specifically, this example supports the following two quantization data types, where AIMQIC can be represented by 1 bit:

[0177] 1: Int8 (QuantType.QInt8): signed 8-bit integer

[0178] 0: UInt8 (QuantType.QUInt8): unsigned 8-bit integer

[0179] Other data types or more data types may also be supported, and accordingly, AIMQIC may also be represented using more bits.

[0180] At step S1130 , the NW parses the received AIMQIC to determine the target quantization precision: either Int8 or UInt8 . It then checks whether an encoder model that meets the AIMQIC precision type requirements exists. If so, it directly executes step S1140 . If no encoder model of the target precision type is found, it performs encoder model quantization, converting the NW encoder model to the target precision type.

[0181] The specific quantization algorithm is not limited in the embodiments of the present application, and a suitable numerical quantization algorithm can be selected according to needs.

[0182] S1140, the NW sends the encoder model of the target precision type to the UE.

[0183] At S1150 , the UE receives the encoder model that meets the AIMQIC requirements in S1120 and completes the deployment and subsequent inference tasks.

[0184] During AI model transmission, the UE instructs the NW to perform AI model quantization by reporting quantization indication configurations. The quantized AI model size is reduced exponentially, significantly reducing air interface transmission overhead, hardware memory and storage usage, lowering device power consumption, and improving model inference efficiency.

[0185] Example 1-2: UE-side training, with separate deployments of the UE and NW. This example uses AI channel state information feedback as an example.

[0186] The method for indicating AI model quantization proposed in this embodiment is not limited to use when the NW side sends the model to the UE side. It is also applicable when the UE transmits the AI ​​model to the NW, and is compatible with both uplink and downlink AI model transmission. Different from Example 1-1, this example proposes a method for indicating AI model quantization in a channel state feedback solution trained only on the UE side, as shown in Figure 12 below.

[0187] In this example, the CSI model (encoder model and decoder model) can be trained entirely on the UE side. The decoder model needs to be deployed on the NW side, so the UE needs to send the decoder model to the NW side. Before sending the decoder model, it is necessary to report the AI ​​Model Quantization Indication Configuration (AIMQIC) to request a quantized decoder model. The specific process example is as follows:

[0188] S1210: The UE sends an AI model quantization configuration request to the NW via UCI. This step is optional and can also be triggered by CSI services to directly trigger the NW to execute S1220.

[0189] S1220: The NW reports AI model quantization configuration information (AIMQIC) to the UE via DCI. Specifically, this example may support the following two quantization data types, and may also support other data types or more data types:

[0190] Int8 (QuantType.QInt8): signed 8-bit integer

[0191] UInt8 (QuantType.QUInt8): unsigned 8-bit integer

[0192] Unlike Example 1-1, this example divides the model into different parameter types and uses different target quantization data types for each. The neural network model parameters are divided into two parts: activation and weight. This example supports different quantization precisions for the activation and weight parameters of the model. The data formats can be (activation: uint8, weight: uint8), (activation: uint8, weight: int8), etc.

[0193] Based on the above description, AIMQIC can use 2 bits to indicate:

[0194] 00: U8U8 as a shorthand for (activation: uint8, weight: uint8);

[0195] 01: U8S8 as a shorthand for (activation: uint8, weight: int8);

[0196] 10: S8U8 is short for (activation: int8, weight: uint8);

[0197] 11: S8S8 is an abbreviation for (activation: int8, weight: int8).

[0198] AIMQIC is not limited to the above settings and can also use fewer or more bits to represent coarser or finer model quantization accuracy requirements. The value of the above AIMQIC and its corresponding precision type can vary, for example, 01 corresponds to U8U8, 00: U8S8, etc.

[0199] At step S1230, the UE parses the received AIMQIC to determine the target quantization precision: one of U8U8, U8S8, S8U8, and S8S8. It queries for a decoder model that meets the AIMQIC precision type requirements. If one exists, it directly executes step S1240. If no decoder model of the target precision type is found, it performs decoder model quantization and converts the decoder model in the UE to a decoder model of the target precision type.

[0200] The specific quantization algorithm is not limited in the embodiments of the present application, and a suitable numerical quantization algorithm can be selected according to needs.

[0201] S1240, the UE uploads the decoder model of the target accuracy type to the NW.

[0202] At S1250, NW receives the decoder model that meets the AIMQIC requirements in S1220 and completes the deployment and subsequent inference tasks.

[0203] During AI model transmission, the NW instructs the UE to perform AI model quantization by reporting quantization indication configurations. The quantized AI model size is reduced exponentially, significantly reducing air interface transmission overhead, hardware memory and storage usage, lowering device power consumption, and improving model inference efficiency.

[0204] Example 2: Quantizing AI Models in a Unilateral Training and Deployment AI Solution

[0205] Example 2-1: NW training, UE-side deployment. This example uses AI beam management as an example.

[0206] In one example, the method of the embodiment of the present application can be used in the AI ​​beam management process. The AI ​​model quantization workflow in the AI-based beam prediction solution is shown in Figure 13:

[0207] The beam prediction model in this example needs to be trained on the NW side and deployed on the UE side, so the NW needs to send the model to the UE side. Before sending the model, it is necessary to report the AI ​​model quantization indication configuration to request a quantized model. The specific process is as follows:

[0208] S1310: The NW sends an AI model quantization configuration request to the UE via DCI. This step is optional and can also be triggered by the service to directly trigger the NW to execute S1320.

[0209] S1320: Based on the received AI model quantization configuration request, the UE reports the AI ​​model quantization configuration information (AIMQIC) to the NW via UCI. Specifically, unlike Example 1-1, this example can support the following quantization data types, and the model parameters can be divided into weight parameters and activation parameters in a non-granular manner. AIMQIC uses 2 bits to indicate:

[0210] 00: Int8 (QuantType.QInt8): signed 8-bit integer

[0211] 01: UInt8 (QuantType.QUInt8): unsigned 8-bit integer

[0212] 10: Float8 (QuantType.QFloat8): 8-bit single-precision floating point type

[0213] 11: Float16 (QuantType.QFloat16): 16-bit single-precision floating point type

[0214] AIMQIC is not limited to the above settings and can also use fewer or more bits to represent coarser or finer model quantization accuracy requirements. The value of the above AIMQIC and its corresponding precision type can vary, for example, 01 corresponds to Int8, 00: UInt8, etc.

[0215] At step S1330 , the NW parses the received AIMQIC and determines the target quantization precision: one of Int8, UInt8, Float8, and Float16. It then searches for a beam prediction model that meets the AIMQIC precision type requirements. If one exists, it directly executes step S1340 . If no beam prediction model of the target precision type is found, it performs model quantization to convert the existing model to the target precision type.

[0216] The specific quantization algorithm is not limited in the embodiments of the present application, and a suitable numerical quantization algorithm can be selected according to needs.

[0217] S1340, the NW sends the beam prediction model of the target accuracy type to the UE.

[0218] At S1350, the UE receives the beam prediction model that meets the AIMQIC requirements in S1320 and completes the deployment and subsequent reasoning tasks.

[0219] During AI model transmission, the UE instructs the NW to perform AI model quantization by reporting quantization indication configurations. The quantized AI model size is reduced exponentially, significantly reducing air interface transmission overhead, hardware memory and storage usage, lowering device power consumption, and improving model inference efficiency.

[0220] Example 2-2: UE training, NW-side deployment. This example uses AI beam management as an example.

[0221] Different from Example 2-1, this example allows the UE to upload the AI ​​model to the NW, as shown in Figure 14:

[0222] Different from Example 2-1, the UE, as the model training side, can upload the quantized model to the NW. The basic steps S1410 to S1440 are similar to those of Example 2-1, but the execution body is different, and the details are not repeated here.

[0223] In S1420, AIMQIC can simultaneously support finer grouping of model parameters with different quantization precisions. Model parameters are further divided into three parts: weight parameters, bias parameters, and activation parameters, each supporting the following two quantization data types:

[0224] Int8 (QuantType.QInt8): signed 8-bit integer

[0225] UInt8 (QuantType.QUInt8): unsigned 8-bit integer

[0226] Therefore, AIMQIC can use 3 bits to indicate:

[0227] 000: U8U8U8, (activation: uint8, bias: uint8, weight: uint8)

[0228] 001: U8U8S8, (activation: uint8, bias: uint8, weight: int8)

[0229] 010: U8S8U8, (activation: uint8, bias: int8, weight: uint8)

[0230] 011: U8S8S8, (activation: uint8, bias: int8, weight: int8)

[0231] 100: S8U8U8, (activation: int8, bias: uint8, weight: uint8)

[0232] 101: S8U8S8, (activation: int8, bias: uint8, weight: int8)

[0233] 110: S8S8U8, (activation: int8, bias: int8, weight: uint8)

[0234] 111: S8S8S8, (activation: int8, bias: int8, weight: int8)

[0235] AIMQIC is not limited to the above settings and can also use fewer or more bits to represent coarser or finer model quantization accuracy requirements. The value of the above AIMQIC and its corresponding precision type can vary, for example, 01 corresponds to Int8, 00: UInt8, etc.

[0236] Single-side training scenarios can also include NW training and NW-side deployment, or UE training and UE-side deployment. In both cases, model transmission is not involved, model quantization is made on a single side, and there is no interaction of AI model quantization information.

[0237] Example 3: Global AI Model Quantification Indication Solution

[0238] The method for indicating AI model quantization in the embodiment of the present application is not limited to one or more use cases (usecase-specific). For example, the method may include user-specific (UE-specific) AI model quantization indication, and the workflow is shown in Figure 15:

[0239] S1510: The UE reports global model quantization configuration information to the NW.

[0240] S1520: The NW performs model quantization on the model of service 1 according to the quantization configuration information.

[0241] S1530: The NW sends the quantized model of service 1 to the UE.

[0242] S1540: The UE completes the model deployment of service 1 and executes reasoning task 1 of the model of service 1.

[0243] If there are other services, such as service 2, service 3, ..., service N, refer to steps S1520 to S1540.

[0244] In this example, there is no need for a specific AI task to drive the AIMQIC request. Instead, the UE can proactively send AIMQIC to the NW after access. AIMQIC provides quantitative guidance for all AI models that need to be sent to the UE.

[0245] AIMQIC can be reported autonomously through UCI or together with the UE capability indication configuration.

[0246] During AI model transmission, the UE instructs the NW to perform AI model quantization by reporting quantization indication configurations. The quantized AI model size is reduced exponentially, significantly reducing air interface transmission overhead, hardware memory and storage usage, lowering device power consumption, and improving model inference efficiency.

[0247] FIG16 is a schematic block diagram of a first communication device 1600 according to an embodiment of the present application. The first communication device 1600 may include:

[0248] The receiving unit 1601 is configured to receive model quantization configuration information from a second communication device;

[0249] The processing unit 1602 is configured to perform model quantization related operations according to the model quantization configuration information.

[0250] In one embodiment, the model quantization configuration information includes target accuracy information.

[0251] In one embodiment, the target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

[0252] In one embodiment, as shown in FIG17 , the first communication device 1700 further includes:

[0253] The first sending unit 1701 is configured to send the first model and / or parameters of the first model to the second communication device when the first communication device includes the first model corresponding to the target accuracy information. For example, the processing unit may check whether there is a first model that matches the target accuracy information, and if so, instruct the first sending unit 1603 to specify the first model and / or parameters of the first model that match the target accuracy information.

[0254] In one embodiment, the processing unit 1602 is further configured to convert the second model into the first model according to the target accuracy information when the first communication device does not include the first model corresponding to the target accuracy information;

[0255] The first communication device further includes: a second sending unit 1702, configured to send the first model and / or parameters of the first model to the second communication device.

[0256] In one embodiment, the first communication device further includes:

[0257] The third sending unit 1703 is configured to send a model quantization configuration request to the second communication device.

[0258] In one embodiment, the model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy, and weight parameter quantization accuracy.

[0259] In one embodiment, the first model is obtained based on a model trained by the first communication device or a sub-model of the model trained by the first communication device.

[0260] In one implementation, the first communication device is a network device, and the second communication device is a terminal device.

[0261] In one embodiment, the first model is an encoder model of a CSI model, the CSI model is trained on the network device, and is deployed on the network device and / or the terminal device.

[0262] In one embodiment, the first communication device is a terminal device, and the second communication device is a network device.

[0263] In one embodiment, the first model is a decoder model of a CSI model, the CSI model is trained on the terminal device, and is deployed on the network device and / or the terminal device.

[0264] In one embodiment, the first communication device is a first terminal device, and the second communication device is a second terminal device.

[0265] In one embodiment, the first model is an encoder model and / or a decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device.

[0266] The first communication devices 1600 and 1700 of the embodiments of the present application can implement the corresponding functions of the first communication device in the aforementioned method embodiment. The processes, functions, implementation methods and beneficial effects corresponding to the various modules (sub-modules, units or components, etc.) in the first communication device can be found in the corresponding descriptions in the above-mentioned method embodiments, and will not be repeated here. It should be noted that the functions described in the various modules (sub-modules, units or components, etc.) in the first communication device of the embodiment of the application can be implemented by different modules (sub-modules, units or components, etc.) or by the same module (sub-module, unit or component, etc.).

[0267] FIG18 is a schematic block diagram of a second communication device 1800 according to an embodiment of the present application. The second communication device 1800 may include:

[0268] The sending unit 1801 is configured to send model quantization configuration information to a first communication device, where the model quantization configuration information is used to instruct the second communication device to perform model quantization-related operations.

[0269] In one embodiment, the model quantization configuration information includes target accuracy information.

[0270] In one embodiment, the target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

[0271] In one embodiment, as shown in FIG19 , the second communication device 1900 further includes:

[0272] The first receiving unit 1901 is configured to receive a first model corresponding to the target accuracy information sent by the first communication device.

[0273] In one embodiment, the second communication device further includes:

[0274] The second receiving unit 1902 is configured to receive a model quantization configuration request sent by the first communication device.

[0275] In one embodiment, the model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy, and weight parameter quantization accuracy.

[0276] In one embodiment, the first model is obtained based on a model trained by the first communication device or a sub-model of the model trained by the first communication device.

[0277] In one implementation, the first communication device is a network device, and the second communication device is a terminal device.

[0278] In one embodiment, the first model is an encoder model of a CSI model, the CSI model is trained on the network device, and is deployed on the network device and / or the terminal device.

[0279] In one embodiment, the first communication device is a terminal device, and the second communication device is a network device.

[0280] In one embodiment, the first model is a decoder model of a CSI model, the CSI model is trained on the terminal device, and is deployed on the network device and / or the terminal device.

[0281] In one embodiment, the first communication device is a first terminal device, and the second communication device is a second terminal device.

[0282] In one embodiment, the first model is an encoder model and / or a decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device.

[0283] The second communication devices 1800 and 1900 of the embodiments of the present application can implement the corresponding functions of the second communication device in the aforementioned method embodiment. The processes, functions, implementation methods and beneficial effects corresponding to the various modules (sub-modules, units or components, etc.) in the second communication device can be found in the corresponding descriptions in the aforementioned method embodiments, and will not be repeated here. It should be noted that the functions described in the various modules (sub-modules, units or components, etc.) in the second communication device of the embodiment of the application can be implemented by different modules (sub-modules, units or components, etc.) or by the same module (sub-module, unit or component, etc.).

[0284] Figure 20 is a schematic structural diagram of a communication device 2000 according to an embodiment of the present application. The communication device 2000 includes a processor 2010, which can call and run a computer program from a memory to enable the communication device 2000 to implement the method according to the embodiment of the present application.

[0285] In one embodiment, the communication device 2000 may further include a memory 2020. The processor 2010 may call and execute a computer program from the memory 2020 to enable the communication device 2000 to implement the method in the embodiment of the present application.

[0286] The memory 2020 may be a separate device independent of the processor 2010 , or may be integrated into the processor 2010 .

[0287] In one embodiment, the communication device 2000 may further include a transceiver 2030 , and the processor 2010 may control the transceiver 2030 to communicate with other devices. Specifically, the transceiver 2030 may send information or data to other devices, or receive information or data sent by other devices.

[0288] The transceiver 2030 may include a transmitter and a receiver. The transceiver 2030 may further include an antenna, and the number of antennas may be one or more.

[0289] In one embodiment, the communication device 2000 may be the first communication device of the embodiment of the present application, and the communication device 2000 may implement the corresponding processes implemented by the first communication device in each method of the embodiment of the present application. For the sake of brevity, they will not be repeated here.

[0290] In one embodiment, the communication device 2000 may be the second communication device of the embodiment of the present application, and the communication device 2000 may implement the corresponding processes implemented by the second communication device in each method of the embodiment of the present application. For the sake of brevity, they will not be repeated here.

[0291] 21 is a schematic structural diagram of a chip 2100 according to an embodiment of the present application. The chip 2100 includes a processor 2110, which can call and execute a computer program from a memory to implement the method according to the embodiment of the present application.

[0292] In one embodiment, the chip 2100 may further include a memory 2120. The processor 2110 may call and execute a computer program from the memory 2120 to implement the method executed by the first communication device or the second communication device in the embodiment of the present application.

[0293] The memory 2120 may be a separate device independent of the processor 2110 , or may be integrated into the processor 2110 .

[0294] In one embodiment, the chip 2100 may further include an input interface 2130. The processor 2110 may control the input interface 2130 to communicate with other devices or chips, and specifically, may obtain information or data sent by other devices or chips.

[0295] In one embodiment, the chip 2100 may further include an output interface 2140. The processor 2110 may control the output interface 2140 to communicate with other devices or chips, and specifically, may output information or data to other devices or chips.

[0296] In one embodiment, the chip can be applied to the first communication device in the embodiment of the present application, and the chip can implement the corresponding processes implemented by the first communication device in each method of the embodiment of the present application. For the sake of brevity, it will not be repeated here.

[0297] In one embodiment, the chip can be applied to the second communication device in the embodiment of the present application, and the chip can implement the corresponding processes implemented by the second communication device in each method of the embodiment of the present application. For the sake of brevity, it will not be repeated here.

[0298] The chips used in the first communication device and the second communication device may be the same chip or different chips.

[0299] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0300] The processor mentioned above may be a general-purpose processor, a digital signal processor (DSP), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or other programmable logic devices, transistor logic devices, discrete hardware components, etc. The general-purpose processor mentioned above may be a microprocessor or any conventional processor, etc.

[0301] The memory mentioned above may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM).

[0302] It should be understood that the above-mentioned memories are exemplary but not restrictive. For example, the memories in the embodiments of the present application may also be static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM RAM (DR RAM), etc. In other words, the memories in the embodiments of the present application are intended to include, but are not limited to, these and any other suitable types of memories.

[0303] FIG22 is a schematic block diagram of a communication system 2200 according to an embodiment of the present application. The communication system 2200 includes a first communication device 2210 and a second communication device 2220 .

[0304] The first communication device 2210 is configured to receive model quantization configuration information from the second communication device; and perform model quantization related operations according to the model quantization configuration information.

[0305] The second communication device 2220 is configured to send model quantization configuration information to the first communication device, where the model quantization configuration information is used to instruct the second communication device to perform model quantization related operations.

[0306] The first communication device 2210 can be used to implement the corresponding functions implemented by the first communication device in the above method, and the second communication device 2220 can be used to implement the corresponding functions implemented by the second communication device in the above method. For the sake of brevity, they are not described here in detail.

[0307] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function in accordance with the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center by wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode to another website, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrations. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a DVD), or a semiconductor medium (eg, a solid state disk (SSD)).

[0308] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0309] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0310] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A communication method, comprising: The first communication device receives model quantization configuration information from the second communication device; The first communication device performs model quantization related operations according to the model quantization configuration information.

2. The method according to claim 1, wherein: The model quantization configuration information includes target accuracy information.

3. The method according to claim 2, wherein: The target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

4. The method according to claim 2 or 3, wherein: The first communication device performs model quantization related operations according to the model quantization configuration information, including: In a case where the first communication device includes a first model corresponding to the target accuracy information, the first model and / or a parameter of the first model is sent to the second communication device.

5. The method according to any one of claims 2 to 4, wherein: The first communication device performs model quantization related operations according to the model quantization configuration information, including: In a case where the first communication device does not include a first model corresponding to the target accuracy information, converting the second model into the first model according to the target accuracy information; The first communication device sends the first model and / or parameters of the first model to the second communication device.

6. The method according to any one of claims 1 to 5, wherein: The method further comprises: The first communication device sends a model quantization configuration request to the second communication device.

7. The method according to any one of claims 1 to 6, wherein: The model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy and weight parameter quantization accuracy.

8. The method according to any one of claims 1 to 7, wherein: The first model is obtained based on a model trained by the first communication device or a sub-model of the model trained by the first communication device.

9. The method according to any one of claims 1 to 8, wherein: The first communication device is a network device, and the second communication device is a terminal device.

10. The method according to claim 9, wherein: The first model is an encoder model of a CSI model, and the CSI model is trained on the network device and deployed on the network device and / or the terminal device.

11. The method according to any one of claims 1 to 8, wherein: The first communication device is a terminal device, and the second communication device is a network device.

12. The method according to claim 11, wherein: The first model is a decoder model of a CSI model, and the CSI model is trained on the terminal device and deployed on the network device and / or the terminal device.

13. The method according to any one of claims 1 to 8, wherein: The first communication device is a first terminal device, and the second communication device is a second terminal device.

14. The method according to claim 13, wherein: The first model is an encoder model and / or a decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device.

15. A communication method, comprising: The second communication device sends model quantization configuration information to the first communication device, where the model quantization configuration information is used to instruct the second communication device to perform model quantization related operations.

16. The method according to claim 15, wherein: The model quantization configuration information includes target accuracy information.

17. The method according to claim 16, wherein: The target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

18. The method according to claim 16 or 17, wherein: The method further comprises: The second communication device receives a first model corresponding to the target accuracy information sent by the first communication device.

19. The method according to any one of claims 15 to 18, wherein: The method further comprises: The second communication device receives the model quantization configuration request sent by the first communication device.

20. The method according to any one of claims 15 to 19, wherein: The model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy and weight parameter quantization accuracy.

21. The method according to any one of claims 15 to 20, wherein: The first model is obtained based on a model trained by the first communication device or a sub-model of the model trained by the first communication device.

22. The method according to any one of claims 15 to 21, wherein: The first communication device is a network device, and the second communication device is a terminal device.

23. The method according to claim 22, wherein: The first model is an encoder model of a CSI model, and the CSI model is trained on the network device and deployed on the network device and / or the terminal device.

24. The method according to any one of claims 15 to 21, wherein: The first communication device is a terminal device, and the second communication device is a network device.

25. The method according to claim 24, wherein: The first model is a decoder model of a CSI model, and the CSI model is trained on the terminal device and deployed on the network device and / or the terminal device.

26. The method according to any one of claims 15 to 21, wherein: The first communication device is a first terminal device, and the second communication device is a second terminal device.

27. The method according to claim 26, wherein: The first model is an encoder model and / or a decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device.

28. A first communication device, comprising: A receiving unit, configured to receive model quantization configuration information from a second communication device; A processing unit is used to perform model quantization related operations according to the model quantization configuration information.

29. The first communication device according to claim 28, wherein: The model quantization configuration information includes target accuracy information.

30. The first communication device according to claim 29, wherein: The target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

31. The first communication device according to claim 29 or 30, wherein: The first communication device comprises: A first sending unit is used to send the first model and / or parameters of the first model to the second communication device when the first communication device includes a first model corresponding to the target accuracy information.

32. The first communication device according to any one of claims 29 to 31, wherein: The processing unit is further configured to convert the second model into the first model according to the target accuracy information when the first communication device does not include the first model corresponding to the target accuracy information; The first communication device further includes: a second sending unit, configured to send the first model and / or parameters of the first model to the second communication device.

33. The first communication device according to any one of claims 28 to 32, wherein: The first communication device further includes: The third sending unit is used to send a model quantization configuration request to the second communication device.

34. The first communication device according to any one of claims 28 to 33, wherein: The model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy and weight parameter quantization accuracy.

35. The first communication device according to any one of claims 28 to 34, wherein: The first model is obtained based on a model trained by the first communication device or a sub-model of the model trained by the first communication device.

36. The first communication device according to any one of claims 28 to 35, wherein: The first communication device is a network device, and the second communication device is a terminal device.

37. The first communication device according to claim 36, wherein: The first model is an encoder model of a CSI model, and the CSI model is trained on the network device and deployed on the network device and / or the terminal device.

38. The first communication device according to any one of claims 28 to 35, wherein: The first communication device is a terminal device, and the second communication device is a network device.

39. The first communication device according to claim 38, wherein: The first model is a decoder model of a CSI model, and the CSI model is trained on the terminal device and deployed on the network device and / or the terminal device.

40. The first communication device according to any one of claims 28 to 35, wherein: The first communication device is a first terminal device, and the second communication device is a second terminal device.

41. The first communication device according to claim 40, wherein: The first model is an encoder model and / or a decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device.

42. A second communication device, comprising: A sending unit is used to send model quantization configuration information to the first communication device, where the model quantization configuration information is used to instruct the second communication device to perform model quantization related operations.

43. The second communication device according to claim 42, wherein: The model quantization configuration information includes target accuracy information.

44. The second communication device according to claim 43, wherein: The target accuracy information includes one or more of the following: a target quantization accuracy of an activation parameter, a target quantization accuracy of a bias parameter, and a target quantization accuracy of a weight parameter.

45. The second communication device according to claim 43 or 44, wherein: The second communication device further includes: The first receiving unit is used to receive a first model corresponding to the target accuracy information sent by the first communication device.

46. ​​The second communication device according to any one of claims 42 to 45, wherein: The second communication device further includes: The second receiving unit is used to receive the model quantization configuration request sent by the first communication device.

47. The second communication device according to any one of claims 42 to 46, wherein: The model quantization accuracy supported by the first communication device includes one or more of the following: activation parameter quantization accuracy, bias parameter quantization accuracy and weight parameter quantization accuracy.

48. The second communication device according to any one of claims 42 to 47, wherein: The first model is obtained based on a model trained by the first communication device or a sub-model of the model trained by the first communication device.

49. The second communication device according to any one of claims 42 to 48, wherein: The first communication device is a network device, and the second communication device is a terminal device.

50. The second communication device according to claim 49, wherein: The first model is an encoder model of a CSI model, and the CSI model is trained on the network device and deployed on the network device and / or the terminal device.

51. The second communication device according to any one of claims 42 to 48, wherein: The first communication device is a terminal device, and the second communication device is a network device.

52. The second communication device according to claim 51, wherein: The first model is a decoder model of a CSI model, and the CSI model is trained on the terminal device and deployed on the network device and / or the terminal device.

53. The second communication device according to any one of claims 42 to 48, wherein: The first communication device is a first terminal device, and the second communication device is a second terminal device.

54. The second communication device according to claim 53, wherein: The first model is an encoder model and / or a decoder model of a CSI model, and the CSI model is trained on the first terminal device and deployed on the first terminal device and / or the second terminal device.

55. A communication device comprising: A processor and a memory, the memory being used to store a computer program, the processor being used to call and run the computer program stored in the memory so that the communication device executes the method as claimed in any one of claims 1 to 14 or any one of claims 15 to 27.

56. A chip, comprising: A processor, configured to call and run a computer program from a memory, so that a device equipped with the chip executes a method as claimed in any one of claims 1 to 14 or any one of claims 15 to 27.

57. A computer-readable storage medium for storing a computer program, which, when executed by a device, causes the device to perform the method as claimed in any one of claims 1 to 14 or any one of claims 15 to 27.

58. A computer program product comprising computer program instructions for causing a computer to perform the method of any one of claims 1 to 14 or any one of claims 15 to 27.

59. A computer program causing a computer to perform the method of any one of claims 1 to 14 or any one of claims 15 to 27.

Citation Information

Patent Citations

  • Information reporting method and device, terminal and readable storage medium

    CN115913486A

  • Communication method and device, terminal and network equipment

    CN116017608A

  • Method and apparatus for feeding back channel state

    CN116032419A

  • Method and apparatus for transmitting model data

    WO2022141544A1

  • Model-based determination of feedback information concerning the channel state

    WO2022212253A1