Model transmission method and apparatus, device, medium and program product

By receiving low-precision models from terminal devices and gradually updating them to full precision using updated parameters, the problem of AI/ML model transmission latency is solved, achieving efficient model loading and task execution, and is suitable for various AI/ML models and application scenarios.

WO2026233311A1PCT designated stage Publication Date: 2026-11-12CHINA MOBILE COMM LTD RES INST +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2026/094892
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-05-06
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

In mobile applications of artificial intelligence/machine learning, there are significant latency issues during model transmission, loading, and execution, especially when terminal device bandwidth is limited, and existing technologies have failed to effectively solve this problem.

Method used

By receiving low-precision models sent by network-side devices and receiving updated parameters while executing tasks, the model is gradually updated to full precision using these parameters, thereby synchronizing the model's difference information and reducing the transmission requirements of full-precision models.

Benefits of technology

It significantly reduces model transmission and loading latency while maintaining the accuracy and flexibility of task execution, making it suitable for a variety of AI/ML models and application scenarios, such as real-time video analytics and autonomous driving.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2026094892_12112026_PF_FP_ABST
    Figure CN2026094892_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The present disclosure relates to a model transmission method and apparatus, a device, a medium and a program product, so as to solve the problem of significant delay generated during transmission, loading and execution of related models. The method comprises: receiving a first model sent by a network side device, and executing a first task on the basis of the first model; when the first task is executed on the basis of the first model, receiving an update parameter sent by the network side device, the update parameter being used for indicating difference information between the first model and a full-precision model; and executing the first task on the basis of the update parameter and the first model, thereby improving the accuracy of task execution. The present disclosure can reduce the delay of model transmission as well as model loading and operating by terminals, while ensuring that the accuracy of task execution is not significantly affected.
Need to check novelty before this filing date? Find Prior Art

Description

Model transmission methods, devices, equipment, media and program products

[0001] Cross-references to related applications

[0002] This disclosure is based on and claims priority to Chinese Patent Application No. 202510589354.2, filed on May 8, 2025, the entire contents of which are hereby incorporated herein by reference. Technical Field

[0003] This disclosure relates to the field of artificial intelligence technology, and in particular to a model transmission method, apparatus, device, medium, and program product. Background Technology

[0004] In mobile applications of Artificial Intelligence (AI) / Machine Learning (ML), such as autonomous driving and real-time video surveillance, pre-trained AI / ML models are used to identify and label objects in images and videos. These applications require terminal devices to run the models with extremely low latency; however, the model communication bandwidth on these devices is also very limited, and even the bandwidth from the server to the terminal is finite. Although the Moving Pictures Experts Group (MPEG) is researching AI / ML model compression methods, such as the next-generation video codec standard MPEG Machine Vision Coding (VCM), AI / ML models can still be very large after compression, leading to long transmission times and high bandwidth consumption, resulting in significant latency during model transmission, loading, and execution. Summary of the Invention

[0005] This disclosure provides a model transmission method, apparatus, device, medium, and program product to solve the problem of significant delays during model transmission, loading, and execution.

[0006] To solve the above-mentioned technical problems, this disclosure is implemented as follows:

[0007] In a first aspect, embodiments of this disclosure provide a model transmission method applied to a terminal, the method comprising:

[0008] Receive the first model sent by the network-side device, and perform the first task based on the first model;

[0009] When performing the first task based on the first model, the update parameters sent by the network-side device are received, the update parameters being used to indicate the difference information between the first model and the full-precision model;

[0010] Execute the first task based on the updated parameters and the first model;

[0011] The execution of the first task based on the updated parameters and the first model includes:

[0012] The first model is updated based on the updated parameters to obtain a full-precision model;

[0013] The first task is performed based on the full-precision model.

[0014] In some embodiments, performing the first task based on the updated parameters and the first model includes:

[0015] Based on the update parameters, at least one node of the first model is updated to obtain a node subset, which includes the updated at least one node;

[0016] The node subset is loaded into memory, and the first task is executed based on the node subset. The process then returns to the step of updating at least one node of the first model based on the update parameters, until all nodes of the first model have been updated.

[0017] In some embodiments, the terminal is configured to support mixed-precision operation.

[0018] In some embodiments, the updating of the first model is performed at least partially in parallel with the execution of the first task.

[0019] In some embodiments, prior to receiving the first model sent by the network-side device, the method further includes:

[0020] Receive target data, wherein the target data is the processing object of the first task;

[0021] The target data is transmitted to the caching module for caching.

[0022] In some embodiments, the method further includes:

[0023] Send a first request to the network-side device, the first request being used to request the acquisition of a full-precision model;

[0024] Receive the full-precision model sent by the network-side device;

[0025] The second task is performed based on the full-precision model.

[0026] In some embodiments, the first model is a model that uses a first data precision to represent model parameters and perform calculations;

[0027] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than the first precision.

[0028] Secondly, embodiments of this disclosure provide a model transmission method applied to a network-side device, the method comprising:

[0029] Send the first model to the terminal;

[0030] Upon receiving the first model at the terminal, update parameters are sent to the terminal, which are used to indicate the difference information between the first model and the full-precision model.

[0031] In some embodiments, before sending the first model to the terminal, the method further includes:

[0032] The initial model is quantized to obtain the full-precision model corresponding to the initial model and the first model.

[0033] In some embodiments, the method further includes:

[0034] Receive a first request sent by the terminal, the first request being used to request the acquisition of a full-precision model;

[0035] The full-precision model is sent to the terminal based on the first request.

[0036] In some embodiments, sending the full-precision model to the terminal based on the first request includes: the sending of the full-precision model and the terminal's task execution using the first model are performed at least in parallel.

[0037] In some embodiments, the first model is a model that uses data representing model parameters and performing calculations with a first data precision, wherein the first data precision is lower than a first threshold.

[0038] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0039] Thirdly, embodiments of this disclosure provide a model transmission device applied to a terminal, the device comprising:

[0040] The first receiving module is configured to receive a first model sent by the network-side device and perform a first task based on the first model;

[0041] The second receiving module is configured to receive update parameters sent by the network-side device when performing the first task based on the first model. The update parameters are used to indicate the difference information between the first model and the full-precision model.

[0042] A first execution module is configured to execute the first task based on the updated parameters and the first model;

[0043] The first execution module includes:

[0044] The first update unit is used to update the first model based on the update parameters to obtain a full-precision model;

[0045] The first execution unit is used to execute the first task based on the full-precision model.

[0046] In some embodiments, the first execution module includes:

[0047] The second update unit is configured to update at least one node of the first model based on the update parameters to obtain a node subset, the node subset including the updated at least one node;

[0048] The second execution unit is configured to load the subset of nodes into memory, execute the first task based on the subset of nodes, and return to the step of updating at least one node of the first model based on the update parameters, until all nodes of the first model have been updated.

[0049] In some embodiments, the apparatus further includes:

[0050] The third receiving module is configured to receive target data, which is the processing object of the first task;

[0051] The transmission module is configured to transmit the target data to the cache module for caching.

[0052] In some embodiments, the apparatus further includes:

[0053] The third sending module is configured to send a first request to the network-side device, the first request being used to request the acquisition of a full-precision model;

[0054] The fourth receiving module is configured to receive the full-precision model sent by the network-side device;

[0055] The second execution module is configured to perform a second task based on the full-precision model.

[0056] In some embodiments, the first model is a model that uses a first data precision to represent model parameters and perform calculations, wherein the first data precision is lower than a first threshold.

[0057] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0058] Fourthly, embodiments of this disclosure provide a model transmission apparatus applied to a network-side device, the apparatus comprising:

[0059] The first sending module is configured to send the first model to the terminal;

[0060] The second sending module is configured to send update parameters to the terminal when the terminal receives the first model. The update parameters are used to indicate the difference information between the first model and the full-precision model.

[0061] In some embodiments, the device is further configured to:

[0062] The initial model is quantized to obtain the full-precision model corresponding to the initial model and the first model.

[0063] In some embodiments, the apparatus further includes:

[0064] The fifth receiving module is configured to receive a first request sent by the terminal, the first request being used to request the acquisition of a full-precision model;

[0065] The fourth sending module is configured to send the full-precision model to the terminal based on the first request.

[0066] In some embodiments, the first model is a model that uses data representing model parameters and performing calculations with a first data precision, wherein the first data precision is lower than a first threshold.

[0067] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0068] Fifthly, embodiments of this disclosure provide a terminal, including a transceiver and a processor, wherein the transceiver is configured to:

[0069] Receive the first model sent by the network-side device, and perform the first task based on the first model;

[0070] When performing the first task based on the first model, the update parameters sent by the network-side device are received, the update parameters being used to indicate the difference information between the first model and the full-precision model;

[0071] The processor is configured to execute the first task based on the updated parameters and the first model;

[0072] The processor is used for:

[0073] The first model is updated based on the updated parameters to obtain a full-precision model;

[0074] The first task is performed based on the full-precision model.

[0075] In some embodiments, the processor is configured to:

[0076] Based on the update parameters, at least one node of the first model is updated to obtain a node subset, which includes the updated at least one node;

[0077] The node subset is loaded into memory, and the first task is executed based on the node subset. The process then returns to the step of updating at least one node of the first model based on the update parameters, until all nodes of the first model have been updated.

[0078] In some embodiments, the transceiver is further configured to:

[0079] Receive target data, wherein the target data is the processing object of the first task;

[0080] The processor is also configured to transmit the target data to a cache module for caching.

[0081] In some embodiments, the transceiver is further configured to:

[0082] Send a first request to the network-side device, the first request being used to request the acquisition of a full-precision model;

[0083] Receive the full-precision model sent by the network-side device;

[0084] The processor is also configured to perform a second task based on the full-precision model.

[0085] In some embodiments, the first model is a model that uses a first data precision to represent model parameters and perform calculations, wherein the first data precision is lower than a first threshold.

[0086] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0087] Sixthly, embodiments of this disclosure provide a network-side device, including a transceiver, the transceiver being configured to:

[0088] Send the first model to the terminal;

[0089] Upon receiving the first model at the terminal, update parameters are sent to the terminal, which are used to indicate the difference information between the first model and the full-precision model.

[0090] In some embodiments, the network-side device further includes a processor, the processor being configured to:

[0091] The initial model is quantized to obtain the full-precision model corresponding to the initial model and the first model.

[0092] In some embodiments, the transceiver is further configured to:

[0093] Receive a first request sent by the terminal, the first request being used to request the acquisition of a full-precision model;

[0094] The full-precision model is sent to the terminal based on the first request.

[0095] In some embodiments, the transmission of the full-precision model and the task execution of the terminal using the first model are at least partially paralleled.

[0096] In some embodiments, the first model is a model that uses data representing model parameters and performing calculations with a first data precision, wherein the first data precision is lower than a first threshold.

[0097] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0098] In a seventh aspect, embodiments of this disclosure provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the model transfer method as described in the first aspect above; or, when the program is executed by the processor, it implements the steps of the model transfer method as described in the second aspect above.

[0099] Eighthly, embodiments of this disclosure provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the model transfer method as described in the first aspect above; or, when executed by a processor, the computer program implements the steps of the model transfer method as described in the second aspect above.

[0100] In a ninth aspect, embodiments of this disclosure provide a computer program product, including computer instructions that, when executed by a processor, implement the steps of the model transfer method as described in the first aspect above; or, when executed by a processor, the computer instructions implement the steps of the model transfer method as described in the second aspect above. Attached Figure Description

[0101] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments of this disclosure will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0102] Figure 1 is a flowchart of one of the model transfer methods provided in this disclosure;

[0103] Figure 2 is an architecture diagram of a model transmission system provided in an embodiment of this disclosure;

[0104] Figure 3 is a schematic diagram of an incremental transmission model provided in an embodiment of this disclosure;

[0105] Figure 4 is a schematic diagram of a full transmission model provided in an embodiment of this disclosure;

[0106] Figure 5 is a second flowchart of a model transfer method provided in an embodiment of this disclosure;

[0107] Figure 6 is a schematic diagram of one of the structures of a model transmission device provided in an embodiment of this disclosure;

[0108] Figure 7 is a second schematic diagram of the structure of a model transmission device provided in an embodiment of this disclosure;

[0109] Figure 8 is a schematic diagram of the structure of a terminal provided in an embodiment of this disclosure;

[0110] Figure 9 is a schematic diagram of the structure of a network-side device provided in an embodiment of this disclosure. Embodiments of the present invention

[0111] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0112] For ease of understanding, the following describes some aspects related to the embodiments of this disclosure:

[0113] In related technologies, AI / ML models used for real-time object recognition, such as the YOLO series or models combining EfficientNet and EfficientDet, may have 50 to 100 million parameters. Alternatively, Transformer models commonly used in vision and speech applications can range in size from several gigabytes (GB) to hundreds of GB, depending on the architecture, model depth, and parameters used. These models may require significant download time and bandwidth, resulting in high startup latency on user equipment (UE). However, in many latency-sensitive applications, the ability to start tasks quickly with lower latency is more important than the accuracy of the results. Therefore, in addition to compressing AI / ML models, the 3rd Generation Partnership Project (3GPP) has also discussed reducing runtime latency by splitting the AI / ML model between the terminal and the network; however, the feasibility of this approach is still under investigation.

[0114] In 3GPP Technical Report (TR) 22.874, the transmission and sharing of AI / ML models has been identified as one of the three key operations for AI / ML-related services. This is because terminals may require a large number of AI / ML models to respond to different tasks and environments. From the perspective of media-based AI / ML use cases defined by 3GPP, the following scenarios require the transmission of AI / ML models: To adapt to background noise or lyrics recognition, specialized AI / ML models for Automatic Speech Recognition (ASR) need to be transmitted to the terminal for device inference; for Natural Language Processing (NLP) of speech: the initial ASR model needs to be downloaded to the UE; then, updated model data needs to be frequently shared with other UEs for distributed / joint learning.

[0115] Related technologies provide an information processing method, including determining channel-related information of an uplink reference signal transmitted by a second communication device, receiving location-related intermediate quantities transmitted by a third communication device, and determining a first dataset for performing a first operation on a model. This method mainly focuses on optimizing the input and label matching of an AI / ML model through channel-related information and location-related intermediate quantities. Related technologies also provide an information transmission method, wherein the information transmission mainly involves terminal movement trajectory-related information, including predicted and actual trajectories, as well as node information, cell identifiers, dwell times, etc., related to these trajectories. A multi-hop transmission mechanism based on predicted and actual trajectories is proposed, including transmitting information between source and target nodes through Xn or NG interfaces. Related technologies also provide a novel process between a server and a distributed client for AI / ML joint learning, mainly applied to federated learning scenarios in 5G networks. This involves distributed learning between user equipment and network equipment using a partially trained AI model. This scheme primarily targets the federated learning process in 5G networks, focusing on distributed training and updating of the model.

[0116] Therefore, none of the aforementioned related technologies have solved the problem of significant delays in the transmission, loading, and execution of related models using the incremental transmission method proposed in this disclosure.

[0117] This disclosure proposes a model transmission method, apparatus, device, medium, and program product to address the significant delays that occur during the transmission, loading, and execution of related models.

[0118] Referring to Figure 1, which is a flowchart of one of the model transmission methods provided in this disclosure, applied to a terminal, the method includes the following steps:

[0119] Step 101: Receive the first model sent by the network-side device, and perform the first task based on the first model.

[0120] In this step, the aforementioned terminal can be a user-side device that interacts with network-side equipment and is configured to receive, send, or process data, such as a mobile phone, computer, or IoT device. The aforementioned network-side equipment can be devices deployed in the network infrastructure that provide access, control, or data processing services to the terminal, such as base stations, access points, core network equipment, and servers.

[0121] The first model mentioned above can be an AI model or ML model configured to perform the first task. The first task can be an application objective executed by the terminal, such as object recognition use cases in various images and videos, such as real-time video analysis and autonomous driving. The first model mentioned above can be a low-precision model, which can refer to a model using numerical precision lower than 32-bit floating-point numbers (FP32), such as 16-bit floating-point (FP16), 8-bit integers (INT8), or mixed precision. It is understood that low-precision models are smaller in size and have faster transmission speeds.

[0122] In some embodiments, before sending the first model, the network-side device can use compression techniques (such as quantization, entropy coding, etc.) to further compress the first model in order to reduce transmission bandwidth consumption. The terminal needs to have the ability to decompress the model (such as entropy decoding) and decompress the compressed data sent by the network-side device to obtain the first model.

[0123] It should be noted that the terminal can start inference immediately after receiving the first model, that is, execute the first task.

[0124] In some embodiments, before receiving the first model sent by the network-side device, the terminal may send a model request to the network-side device, the model request being used to request the acquisition of the first model.

[0125] Step 102: When performing the first task based on the first model, receive update parameters sent by the network-side device, wherein the update parameters are used to indicate the difference information between the first model and the full-precision model.

[0126] In this step, the updated parameters can be obtained by analyzing the differences in parameters and structure between the full-precision model and the low-precision model, extracting key information about changes caused by the reduction in precision. This difference information may include changes in model weight values, adjustments to certain parameters, etc. The full-precision model mentioned above can be a model that uses 32-bit floating-point numbers (FP32) for calculation, retaining high numerical precision, and can also be called a high-precision model.

[0127] It is understandable that the terminal continuously runs the first model and processes the first task, and receives updated parameters sent by the network-side device while performing the first task.

[0128] Step 103: Execute the first task based on the updated parameters and the first model.

[0129] In this step, the above-mentioned execution of the first task based on the updated parameters and the first model can be achieved by the terminal updating the first model based on the updated parameters and then using the updated first model to execute the first task.

[0130] The execution of the first task based on the updated parameters and the first model includes:

[0131] The first model is updated based on the updated parameters to obtain a full-precision model;

[0132] The first task is performed based on the full-precision model.

[0133] In some embodiments, updating the first model based on the update parameters to obtain a full-precision model can be achieved by the terminal loading the updated model all at once using a caching mechanism when it receives the update parameters. That is, all nodes load the updated model uniformly after the update, so that all nodes can use the same version of the model synchronously. This can avoid result conflicts caused by some nodes not being updated, thereby improving the accuracy of task execution.

[0134] In this embodiment, a first model sent by a network-side device is received, and a first task is executed based on the first model. When executing the first task based on the first model, update parameters sent by the network-side device are received, the update parameters indicating the difference between the first model and the full-precision model. The first task is executed based on the update parameters and the first model. Executing the first task based on the update parameters and the first model includes: updating the first model based on the update parameters to obtain a full-precision model; and executing the first task based on the full-precision model. Thus, through the above steps, the terminal can immediately start running the first model and executing the first task after receiving the first model, and update the first model based on the update parameters while executing the first task to gradually approach the performance of the full-precision model. This eliminates the need to retransmit the full-precision model, reducing bandwidth consumption and improving the model accuracy and effectiveness of the first task. Overall, this significantly reduces the latency of model transmission and terminal loading and running of the model, while ensuring that the accuracy of task execution is not significantly affected.

[0135] It should be noted that the above method is applicable to a variety of AI / ML models (such as YOLO, Transformer, etc.) and can be applied to a variety of application scenarios, such as real-time video analysis, autonomous driving, and mobile device applications. It has wide applicability and flexibility. For real-time video analysis, the terminal starts video analysis immediately after receiving the first model and updates the parameters to improve the analysis accuracy. For autonomous driving, the terminal starts environmental perception immediately after receiving the low-precision model and updates the parameters to improve the perception accuracy.

[0136] In some embodiments, performing the first task based on the updated parameters and the first model includes:

[0137] Based on the update parameters, at least one node of the first model is updated to obtain a node subset, which includes the updated at least one node;

[0138] The node subset is loaded into memory, and the first task is executed based on the node subset. The process then returns to the step of updating at least one node of the first model based on the update parameters, until all nodes of the first model have been updated.

[0139] In some embodiments, the nodes in the first model can be updated in batches at preset time intervals until all nodes have been updated. Loading a subset of nodes into memory can be achieved through a caching mechanism, where the updated subset of nodes can be loaded into memory temporarily and released after the operation is completed.

[0140] Understandably, when the terminal supports mixed-precision operations, the updated model can be loaded incrementally, meaning updates are applied only to a subset of nodes, and these updated subsets of nodes are loaded into memory through a caching mechanism. In some embodiments, the updating of the first model and the execution of the first task are performed at least partially in parallel.

[0141] In this implementation, by loading the updated subset of nodes sequentially, the delay in executing tasks during the model update process can be reduced, thereby improving the stability of terminal task execution.

[0142] In some embodiments, prior to receiving the first model sent by the network-side device, the method further includes:

[0143] Receive target data, wherein the target data is the processing object of the first task;

[0144] The target data is transmitted to the caching module for caching.

[0145] In some embodiments, the target data mentioned above can be input data processed by the first task, such as video frames in a video surveillance task, images in an image classification task, and audio files in a speech recognition task. The caching module mentioned above can be a temporary storage area in the terminal, configured to temporarily store data for later fast access.

[0146] In this embodiment, by transmitting the target data to the caching module for caching, data loss can be avoided if the first model is not ready, thus ensuring that the continuity of task execution is not affected and improving the accuracy of task execution.

[0147] For example, Figure 2 is an architecture diagram of a model transmission system provided in an embodiment of this disclosure. As shown in Figure 2, the model transmission system includes a terminal and a network-side application, i.e., a network-side device. The process is as follows:

[0148] 1. The terminal requests the network-side device to send an AI model, the network-side device sends a low-precision model to the terminal, and the terminal immediately executes the task;

[0149] 2. The network-side device asynchronously sends AI model difference packets to the terminal, and the terminal cache is updated. This mechanism reduces initial latency and improves accuracy through difference packets.

[0150] 3. Terminal caching mechanism, supporting the following two modes:

[0151] (1) Full load at once: directly replace with the full-precision model;

[0152] (2) Mixed precision incremental loading: Only some nodes are updated;

[0153] 4. Input buffer: In scenarios such as video surveillance, the terminal briefly buffers input data to avoid frame loss affecting continuity;

[0154] 5. Asynchronous Updates: Model transmission and task execution are performed in parallel, and update loading does not interrupt the inference process. In some embodiments, the transmission of the full-precision model and the task execution using the first model by the terminal are performed at least partially in parallel.

[0155] In some embodiments, the method further includes:

[0156] Send a first request to the network-side device, the first request being used to request the acquisition of a full-precision model;

[0157] Receive the full-precision model sent by the network-side device;

[0158] The second task is performed based on the full-precision model.

[0159] Understandably, a terminal can request the network-side device to directly transmit the full-precision model or incremental transmission based on its own capabilities and network conditions. Incremental transmission means that the network-side device first transmits the low-precision model, and then transmits updated parameters to reconstruct the high-precision model.

[0160] For example, when network conditions are poor, a low-precision model, i.e., the first model, can be selected to reduce latency, while when network conditions are good, a full-precision model can be selected to improve accuracy.

[0161] In this implementation, the terminal can select an appropriate model accuracy according to actual needs, thereby finding the best balance between accuracy and transmission time, which can improve the flexibility and adaptability of model transmission.

[0162] In some embodiments, the first model is a model that uses a first data precision to represent model parameters and perform calculations, wherein the first data precision is lower than a first threshold.

[0163] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0164] In some embodiments, the first model mentioned above may refer to a low-precision model. A low-precision model is a model that uses lower data precision in representing model parameters and in the calculation process. It may refer to a model that uses numerical precision lower than FP32. Common data precision types include half-precision floating-point numbers (FP16), integers (such as INT8), etc.

[0165] Understandably, these low-precision data types consume less memory during training and inference, and can accelerate computation on specific hardware, such as Graphics Processing Units (GPUs) that support low-precision calculations. Using low-precision models can effectively reduce model storage size, decrease model transfer and loading time, and improve computational efficiency, especially suitable for resource-constrained environments such as mobile devices and edge computing devices. However, due to the reduced data precision, low-precision models may sacrifice some accuracy and performance, resulting in a certain loss of precision.

[0166] The aforementioned full-precision model refers to a model that uses high data precision to represent model parameters and perform calculations. The most common is single-precision floating-point numbers (FP32), and in some scenarios with extremely high precision requirements, even double-precision floating-point numbers (FP64) may be used.

[0167] Understandably, full-precision models can more accurately represent the model's parameters and computational results during training and inference, thus typically achieving higher accuracy and better performance. However, full-precision models consume more memory, have higher storage and transmission costs, and are relatively slower in computation, especially in large-scale models and complex computational tasks, placing higher demands on hardware resources.

[0168] In this embodiment, the first model is a model that uses a first data precision to represent model parameters and perform calculations. The first data precision is lower than a first threshold. By transmitting the first model, the transmission time of the model can be significantly reduced, thereby reducing the startup delay of inference.

[0169] For example, an embodiment is used to compare two schemes: full-precision transmission in related technologies and incremental transmission proposed in this disclosure. The comparison is mainly based on the following indicators and parameters:

[0170] Model accuracy: The accuracy of the decoded and reconstructed model. If it is a model update, it is the accuracy of the low-precision model after the decoded model update.

[0171] Startup latency: The time it takes for the terminal to begin inference, measured in seconds. It is the sum of encoding time, decoding time, and reconstruction time.

[0172] Model size: The size of the model transmitted from the network-side device to the terminal. If the model is encoded when it is sent, it is the size of the compressed bitstream.

[0173] Figure 3 is a schematic diagram of an incremental transmission model provided by an embodiment of this disclosure. As shown in Figure 3, the server first sends a low-precision version of the model (at time t) to the terminal. After the terminal receives the model (after time t1), it begins to deploy the model in the current task. Later, the server sends an updated model (containing the differences between the full-precision model and the low-precision model) to the terminal. After the terminal fully receives the model (after time t2), it begins to deploy the updated full-precision model. Figure 4 is a schematic diagram of a full transmission model provided by an embodiment of this disclosure. As shown in Figure 4, the server first sends a full-precision version of the model (at time t) to the terminal. After the terminal receives the model (after time t3), it begins to deploy the model in the current task.

[0174] In Figure 3, terminal UE1 requests the bit increment transmission of the full-precision model, while in Figure 4, terminal UE2 requests the direct transmission of the full-precision model. This is considered the anchor point. At the anchor point, the server directly sends the full-precision model to UE2, while for UE1, the server first sends a low-precision version of the model, followed by update parameters. These update parameters contain the differences between the full-precision and low-precision models, and include the following steps:

[0175] Step 1: Load the model and data:

[0176] Example: The following process uses the BIT library to create a pre-trained ImageNet float32 model and loads data for retraining or testing the model.

[0177] Initialize the model by calling the `BIT.create_model_instances()` method and passing in the relevant parameters. These parameters include: `model_arch`: model architecture; `dataset_id`: dataset ID; `dataset_path`: path where the dataset is stored; `lr`: learning rate; `batch_size`: training batch size; `seed`: random seed, possibly used to ensure the reproducibility of experiments; and `num_workers`: number of worker threads used during data loading.

[0178] Step 2: Quantize the loaded model into two different precision models: one low precision (e.g., 16-bit) and one high (full) precision (e.g., 32-bit).

[0179] First, iterate through the list containing model IDs and perform an approx() quantization operation on each model, gradually reducing the precision from float32 or updating to higher bit precision. approx() is likely a quantization function that takes three arguments: Models["approx_info_list"][i].approx_info: approximate information for each model, potentially used to guide the quantization process; Models["models_info"][i]: stores relevant information about the model (potentially including architecture, parameters, etc.); Models["approx_data_list"][i]: stores additional data needed during the quantization process.

[0180] Step 3: Encode (or compress) the quantized model to generate a bitstream and perform calculations:

[0181] a. Calculate the time required to encode each quantization model;

[0182] b. Calculate the size of the bitstream generated by each quantization model.

[0183] `coder.encode()` is the `encode()` method of the encoder, used to encode the quantized model and update information. Input parameters: `enc_info`: Additional information required for encoding, possibly including encoding configuration or hyperparameters; `Models["models_info"][i]`: Basic information about the current model; `approx_data_quant`: The already quantized model data; `dec_approx_param_list[i]`: Parameters required for decoding (possibly used to recover the model); `None`: May represent default parameters or placeholders.

[0184] Step 4: Decode the bitstream:

[0185] a. Calculate the time required to decode each bit stream.

[0186] `coder.decode()` is the encoder's `decode()` method, used to decode the quantized model and update information. It reads the bitstream file: loading the encoded binary data from the specified path. It then uses `coder.decode` to restore the compressed bitstream to approximate model data. Key parameters include model metadata (`dec_model_info`) and decoding configuration (`dec_approx_param_list[i]`). `hls_bytes` may be used to record intermediate results of layered encoding (such as entropy-encoded byte streams). It records decoding time: statistically analyzes the decoding process duration for performance analysis or optimization.

[0187] Step 5: Reconstruct the decoded bitstream, that is, convert the decoded quantization model into a high (full) precision model:

[0188] a. Calculate the time required to reconstruct and decode the high (full) accuracy model.

[0189] The decoded approximation data is converted into usable floating-point model parameters using `approximator.rec`. The parameter list defined in the model definition is traversed to ensure the decoded parameters are complete, with missing items initialized to zero to avoid model inference anomalies due to missing parameters. The operation type is set to "Incremental Update Mode" (id == "model.update"): merging the decoded parameters with global parameters by module to generate a new state dictionary; or to direct test mode: directly using the decoded parameters for model performance evaluation.

[0190] Step 6: Calculate the evaluation index parameters:

[0191] Call the test_model method to evaluate the accuracy (rec_acc) of the merged model.

[0192] The experimental results are shown in Table 1 below:

[0193] Table 1

[0194]

[0195] Experimental conclusion: The proposed solution can achieve similar accuracy to the direct transmission of high (full) precision models.

[0196] This implementation method can reduce model transmission time while maintaining model performance, thereby improving user experience and application response speed.

[0197] Referring to Figure 5, which is a second flowchart of a model transmission method provided in this embodiment of the present disclosure, applied to a network-side device, the method includes the following steps:

[0198] Step 501: Send the first model to the terminal;

[0199] Step 502: When the terminal receives the first model, send update parameters to the terminal. The update parameters are used to indicate the difference information between the first model and the full-precision model.

[0200] It should be noted that this embodiment is an implementation of the network-side device corresponding to the embodiment shown in Figure 1. The implementation method can be referred to the relevant description in the embodiment shown in Figure 1. To avoid repetition, this embodiment will not be described again.

[0201] In some embodiments, before sending the first model to the terminal, the method further includes:

[0202] The initial model is quantized to obtain the full-precision model corresponding to the initial model and the first model.

[0203] In some embodiments, the initial model mentioned above may refer to the original model configured to perform the first task, which may be a trained AI model or an ML model.

[0204] Understandably, network-side devices can store model versions with different precision, and have both low-precision and high-precision model versions, such as a 16-bit integer low-precision version and a 32-bit floating-point high-precision version model; they can also reduce the model precision and size through quantization techniques, such as converting model parameters from 32-bit floating-point numbers to 16-bit integers.

[0205] In some embodiments, the method further includes:

[0206] Receive a first request sent by the terminal, the first request being used to request the acquisition of a full-precision model;

[0207] The full-precision model is sent to the terminal based on the first request.

[0208] In some embodiments, the first model is a model that uses data representing model parameters and performing calculations with a first data precision, wherein the first data precision is lower than a first threshold.

[0209] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0210] The above implementation methods can be referred to the relevant descriptions in the embodiments shown in Figure 1. To avoid repetition, these descriptions will not be repeated in this embodiment.

[0211] Referring to Figure 6, which is a schematic diagram of one of the structures of a model transmission device provided in this embodiment of the present disclosure, applied to a terminal, as shown in Figure 6, the model transmission device 600 includes:

[0212] The first receiving module 601 is configured to receive a first model sent by a network-side device and perform a first task based on the first model;

[0213] The second receiving module 602 is configured to receive update parameters sent by the network-side device when performing the first task based on the first model. The update parameters are used to indicate the difference information between the first model and the full-precision model.

[0214] The first execution module 603 is configured to execute the first task based on the updated parameters and the first model;

[0215] The first execution module 603 includes:

[0216] The first update unit is configured to update the first model based on the update parameters to obtain a full-precision model;

[0217] The first execution unit is configured to execute the first task based on the full-precision model.

[0218] In some embodiments, the first execution module 603 includes:

[0219] The second update unit is configured to update at least one node of the first model based on the update parameters to obtain a node subset, the node subset including the updated at least one node;

[0220] The second execution unit is configured to load the subset of nodes into memory, execute the first task based on the subset of nodes, and return to the step of updating at least one node of the first model based on the update parameters, until all nodes of the first model have been updated.

[0221] In some embodiments, the apparatus further includes:

[0222] The third receiving module is configured to receive target data, which is the processing object of the first task;

[0223] The transmission module is configured to transmit the target data to the cache module for caching.

[0224] In some embodiments, the apparatus further includes:

[0225] The third sending module is configured to send a first request to the network-side device, the first request being used to request the acquisition of a full-precision model;

[0226] The fourth receiving module is configured to receive the full-precision model sent by the network-side device;

[0227] The second execution module is configured to perform a second task based on the full-precision model.

[0228] In some embodiments, the first model is a model that uses a first data precision to represent model parameters and perform calculations, wherein the first data precision is lower than a first threshold.

[0229] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0230] It should be noted that the model transmission device provided in this embodiment is a device capable of executing the above-described model transmission method. Therefore, all implementations of the above-described model transmission method embodiments are applicable to this device and can achieve the same or similar beneficial effects. To avoid repetition, this embodiment will not elaborate further.

[0231] Referring to Figure 7, which is a second structural schematic diagram of a model transmission device provided in this embodiment of the present disclosure, applied to a network-side device, as shown in Figure 7, the model transmission device 700 includes:

[0232] The first sending module 701 is configured to send the first model to the terminal;

[0233] The second sending module 702 is configured to send update parameters to the terminal when the terminal receives the first model. The update parameters are used to indicate the difference information between the first model and the full-precision model.

[0234] In some embodiments, the device is further configured to:

[0235] The initial model is quantized to obtain the full-precision model corresponding to the initial model and the first model.

[0236] In some embodiments, the apparatus further includes:

[0237] The fifth receiving module is configured to receive a first request sent by the terminal, the first request being used to request the acquisition of a full-precision model;

[0238] The fourth sending module is configured to send the full-precision model to the terminal based on the first request.

[0239] In some embodiments, the first model is a model that uses data representing model parameters and performing calculations with a first data precision, wherein the first data precision is lower than a first threshold.

[0240] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0241] It should be noted that the model transmission device provided in this embodiment is a device capable of executing the above-described model transmission method. Therefore, all implementations of the above-described model transmission method embodiments are applicable to this device and can achieve the same or similar beneficial effects. To avoid repetition, this embodiment will not elaborate further.

[0242] Referring to Figure 8, this embodiment of the present disclosure also provides a terminal, including a bus 801, a transceiver 802, an antenna 803, a bus interface 804, a processor 805, and a memory 806.

[0243] Transceiver 802 is configured as follows:

[0244] Receive the first model sent by the network-side device, and perform the first task based on the first model;

[0245] When performing the first task based on the first model, the update parameters sent by the network-side device are received, the update parameters being used to indicate the difference information between the first model and the full-precision model;

[0246] Furthermore, the processor 805 is configured to perform the first task based on the updated parameters and the first model;

[0247] The processor 805 is configured to:

[0248] The first model is updated based on the updated parameters to obtain a full-precision model;

[0249] The first task is performed based on the full-precision model.

[0250] In Figure 8, a bus architecture (represented by bus 801) is shown. Bus 801 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 805 and memory represented by memory 806. Bus 801 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 804 provides an interface between bus 801 and transceiver 802. Transceiver 802 may be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 805 is transmitted over a wireless medium via antenna 803, which further receives data and transmits data to processor 805.

[0251] The processor 805 manages the bus 801 and handles general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. The memory 806 can be configured to store data used by the processor 805 during operation.

[0252] In some embodiments, the processor 805 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a complex programmable logic device (CPLD).

[0253] In some embodiments, the processor 805 is configured to:

[0254] Based on the update parameters, at least one node of the first model is updated to obtain a node subset, which includes the updated at least one node;

[0255] The node subset is loaded into memory, and the first task is executed based on the node subset. The process then returns to the step of updating at least one node of the first model based on the update parameters, until all nodes of the first model have been updated.

[0256] In some embodiments, the transceiver 802 is further configured to:

[0257] Receive target data, wherein the target data is the processing object of the first task;

[0258] The processor is also configured to transmit the target data to a cache module for caching.

[0259] In some embodiments, the transceiver 802 is further configured to:

[0260] Send a first request to the network-side device, the first request being used to request the acquisition of a full-precision model;

[0261] Receive the full-precision model sent by the network-side device;

[0262] The processor is also configured to perform a second task based on the full-precision model.

[0263] In some embodiments, the first model is a model that uses a first data precision to represent model parameters and perform calculations, wherein the first data precision is lower than a first threshold.

[0264] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0265] It should be noted that the terminal provided in this embodiment is a device capable of executing the above-described model transmission method. Therefore, all implementations of the above-described model transmission method embodiments are applicable to this electronic device and can achieve the same or similar beneficial effects. To avoid repetition, this embodiment will not elaborate further.

[0266] Referring to Figure 9, this embodiment of the present disclosure also provides a network-side device, including a bus 901, a transceiver 902, an antenna 903, a bus interface 904, a processor 905, and a memory 906.

[0267] Transceiver 902 is configured as follows:

[0268] Send the first model to the terminal;

[0269] Upon receiving the first model at the terminal, update parameters are sent to the terminal, which are used to indicate the difference information between the first model and the full-precision model.

[0270] In Figure 9, a bus architecture (represented by bus 901) is shown. Bus 901 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 905 and memory represented by memory 906. Bus 901 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 904 provides an interface between bus 901 and transceiver 902. Transceiver 902 may be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 905 is transmitted over a wireless medium via antenna 903, which further receives data and transmits it to processor 905.

[0271] Processor 905 manages bus 901 and general processing, and also provides various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. Memory 906 can be configured to store data used by processor 905 during operation.

[0272] In some embodiments, the processor 905 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a complex programmable logic device (CPLD).

[0273] In some embodiments, the processor 905 is configured to:

[0274] The initial model is quantized to obtain the full-precision model corresponding to the initial model and the first model.

[0275] In some embodiments, the transceiver 902 is further configured to:

[0276] Receive a first request sent by the terminal, the first request being used to request the acquisition of a full-precision model;

[0277] The full-precision model is sent to the terminal based on the first request.

[0278] In some embodiments, the first model is a model that uses data representing model parameters and performing calculations with a first data precision, wherein the first data precision is lower than a first threshold.

[0279] The full-precision model is a model that uses data of a second precision to represent model parameters and perform calculations, where the second precision is higher than a second threshold.

[0280] It should be noted that the network-side device provided in this embodiment is an apparatus capable of executing the above-described model transmission method. Therefore, all implementations of the above-described model transmission method embodiments are applicable to this electronic device and can achieve the same or similar beneficial effects. To avoid repetition, this embodiment will not elaborate further.

[0281] This disclosure also provides an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the various processes of the above-described model transmission method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here.

[0282] This disclosure also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described model transmission method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0283] This disclosure also provides a computer program product, including computer instructions. When executed by a processor, these computer instructions implement the various processes of the above-described model transmission method embodiments and achieve the same technical effects. To avoid repetition, they will not be described again here.

[0284] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0285] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this disclosure.

[0286] The embodiments of this disclosure have been described above with reference to the accompanying drawings. However, this disclosure is not limited to the above-described embodiments. The above-described embodiments are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of this disclosure without departing from the spirit and scope of the claims, and all of these modifications are within the protection scope of this disclosure.

Claims

1. A model transmission method applied to a terminal, comprising: receiving a first model sent by a network-side device, and performing a first task based on the first model; in a case where the first task is performed based on the first model, receiving an update parameter sent by the network-side device, the update parameter being used to indicate difference information between the first model and a full-precision model; performing the first task based on the update parameter and the first model; the performing the first task based on the update parameter and the first model comprises: updating the first model based on the update parameter to obtain the full-precision model; performing the first task based on the full-precision model.

2. The method of claim 1, wherein, the performing the first task based on the update parameter and the first model comprises: updating at least one node of the first model based on the update parameter to obtain a node subset, the node subset comprising the updated at least one node; loading the node subset into a memory and performing the first task based on the node subset, and returning to the step of updating at least one node of the first model based on the update parameter until all nodes of the first model are updated. 3.The method of claim 2, wherein the terminal is configured to support a mixed-precision operation.

4. The method of claim 1, wherein, the updating of the first model and the performing of the first task are at least partially parallel.

5. The method of claim 1, wherein, before the receiving the first model sent by the network-side device, the method further comprises: receiving target data, the target data being a processing object of the first task; and transmitting the target data to a cache module for caching.

6. The method of claim 1, wherein, the method further comprises: sending a first request to the network-side device, the first request being used to request to obtain the full-precision model; receiving the full-precision model sent by the network-side device; and performing a second task based on the full-precision model.

7. The method of any one of claims 1-6, wherein, the first model is a model that uses a first data precision to represent model parameters and performs calculations; the full-precision model is a model that uses a second data precision to represent model parameters and performs calculations, the second data precision being higher than the first data precision. 8.A model transmission method applied to a network-side device, comprising: sending a first model to a terminal; in a case where the terminal receives the first model, sending an update parameter to the terminal, the update parameter being used to indicate difference information between the first model and a full-precision model.

9. The method of claim 8, wherein, before the sending the first model to the terminal, the method further comprises: performing quantization processing on an initial model to obtain the full-precision model corresponding to the initial model and the first model.

10. The method of claim 8, wherein, the method further comprises: receiving a first request sent by the terminal, the first request being used to request to obtain the full-precision model; and sending the full-precision model to the terminal based on the first request.

11. The method of claim 10, wherein, the sending the full-precision model to the terminal based on the first request comprises: the sending of the full-precision model and the performing of a task by the terminal using the first model are at least partially parallel.

12. The method of any one of claims 8-11, wherein, The first model is a model that uses data of a first data precision to represent model parameters and perform calculations; The full-precision model is a model that uses data of a second data precision to represent model parameters and perform calculations, the second data precision being higher than the first data precision.

13. A model transmission apparatus applied to a terminal, comprising: a first receiving module configured to receive a first model sent by a network-side device and perform a first task based on the first model; a second receiving module configured to, in a case where the first task is performed based on the first model, receive an update parameter sent by the network-side device, the update parameter being used to indicate difference information between the first model and a full-precision model; a first execution module configured to perform the first task based on the update parameter and the first model; the first execution module comprises: a first update unit configured to update the first model based on the update parameter to obtain the full-precision model; a first execution unit configured to perform the first task based on the full-precision model.

14. A model transmission apparatus applied to a network-side device, comprising: a first sending module configured to send a first model to a terminal; a second sending module configured to, in a case where the first model is received by the terminal, send an update parameter to the terminal, the update parameter being used to indicate difference information between the first model and a full-precision mode.

15. A terminal comprising a transceiver and a processor, wherein the transceiver is configured to: receive a first model sent by a network-side device and perform a first task based on the first model; and in a case where the first task is performed based on the first model, receive an update parameter sent from the network-side device, the update parameter being used to indicate difference information between the first model and the full-precision model. The processor is configured to perform the first task based on the update parameter and the first model. The processor is configured to: update the first model based on the update parameter to obtain the full-precision model; and perform the first task based on the full-precision model.

16. A network-side device comprising a transceiver, wherein the transceiver is configured to: send a first model to a terminal; and in a case where the first model is received by the terminal, send an update parameter to the terminal; the update parameter being used to indicate difference information between the first model and a full-precision model.

17. The network-side device of claim 16, wherein, The transceiver is further configured to: receive a first request sent by the terminal, the first request being used to request the full-precision model; and send the full-precision model to the terminal based on the first request, wherein the sending of the full-precision model is at least partially parallel to the task execution of the terminal using the first model.

18. An electronic device, comprising: A processor, a memory, and a program stored on the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the model transmission method according to any one of claims 1 to 7; or the program, when executed by the processor, implements the steps of the model transmission method according to claim 8 to 12.

19. A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the model transmission method according to any one of claims 1 to 7; or, when executed by a processor, implements the steps of the model transmission method according to any one of claims 8 to 12.

20. A computer program product comprising computer instructions which, when executed by a processor, implement the steps of the model transmission method according to any one of claims 1 to 7; or, when executed by a processor, implement the steps of the model transmission method according to any one of claims 8 to 12.