Methods, apparatuses and systems of model processing
The PSC system efficiently converts large AI models into smaller, customized models for resource-constrained terminals by determining and generating target models aligned with terminal needs, addressing deployment challenges of large AI models.
Patent Information
- Application Number
- PCT/CN2024/090393
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-04-28
- Publication Date
- 2025-11-06
AI Technical Summary
Large AI models, such as large language models (LLM) and generative pre-trained transformers (GPT), are often too heavy or costly for efficient deployment due to redundant knowledge and high storage/computing requirements, necessitating a method to convert them into smaller, customized models for resource-constrained terminals.
A processing service controller (PSC) facilitates the conversion of large pre-trained models into smaller, customized models by determining and generating target models based on terminal requirements, utilizing model generators (MGs) with ample resources, and managing model holders (MHs) to ensure efficient deployment.
Enables efficient and flexible deployment of smaller AI models by aligning them with terminal capabilities, reducing computational and storage demands while maintaining accuracy and complexity requirements.
Smart Images

Figure CN2024090393_06112025_PF_FP_ABST
Abstract
Description
METHODS, APPARATUSES AND SYSTEMS OF MODEL PROCESSINGTECHNICAL FIELD
[0001] The present application relates to the field of wireless communication technologies, and in particular to methods, apparatuses and systems of model processing.BACKGROUND
[0002] Large artificial intelligence (AI) models such as large language model (LLM) and generative pre-trained transformer (GPT) are becoming increasingly popular. A large AI model is often trained to solve one or multiple complicated problems and can be used for multiple applications. Some application is highly specialized such that a large pre-trained AI model may include redundant knowledge (outside the subject domain) and / or is too heavy or costly to be used (e.g., requiring too much storage space and computing power, limiting deployment possibilities) .
[0003] When the large pre-trained AI model (which is also referred to as a reference model) is reused for an application (whether for fine turning or for inferencing) , it is desirable to convert (e.g. by downsizing or compressing) the large pre-trained AI model into a small, customized AI model for the application to improve deployment efficiency and flexibility.
[0004] However, model processing (such as downsizing or compressing) requires massive computing and storage resources, which are just a terminal lacks. How to obtain a processed model for an apparatus lacking computing or storage resources is a problem that need to be solved.SUMMARY
[0005] Embodiments of the present application provides various methods, apparatuses and systems of model processing, in order to enable an apparatus lacking computing or storage resources to obtain a processed model.
[0006] In a first aspect, an embodiment of the present application provides a method of model processing, the implementer of the method, i.e., a fourth apparatus, may be a server or a chip applied to a server, the fourth apparatus may also be referred to as a processing service controller (PSC) . In the following description, a PSC is an example of the fourth apparatus, not a restriction. The method includes:
[0007] receiving first information from a first apparatus, wherein the first information indicates a first model (also referred to as the reference model hereinafter) ;
[0008] determining model information, according to the first information, the model information is associated with a second model (also referred to as the target model hereinafter) ;
[0009] sending second information to a second apparatus, wherein the second information includes at least part of the model information.
[0010] In this embodiment, the first information may be, for example, a reference model ID included in an AF request (step 1 in Fig. 2) . The second information may be, for example, configuration information used for configuring (steps 5a-5c in Fig. 2) the second apparatus to generate the second model. By receiving the first information, a PSC could determine the unprocessed model required by a terminal (an example of the first apparatus) , i.e., the first model, and then determine the processed model corresponding to the first model, i.e., the second model. The PSC could send second information to the second apparatus, to instruct the second apparatus process the first model, the second apparatus may be a model generator (MG) with enough computing and storage resources to process the first model, so that the terminal or the AF could obtain a processed model in case of lacking computing or storage resources.
[0011] In an optional embodiment of the first aspect, the second model is smaller than the first model.
[0012] “The second model is smaller than the first model” may refer to that, the number of parameters of the second model is less than the number of parameters of the first model, and / or, the structure of the second model is simpler than that of the first model. By providing a smaller model to a terminal, the deployment efficiency and flexibility for the second model could be improved.
[0013] In an optional embodiment of the first aspect, the model information includes at least one of the following:
[0014] an identifier (ID) of the second model, complexity information of the second model, accuracy information of the second model, route information for obtaining the second model, structure information of the second model, information of specifying value of model parameters of the second model, information of one or more datasets to be used for training the second model.
[0015] The second model may be a variant of the first model or correspond to a version of the first model, and the ID of the second model indicates or identifies the variant of the first model or the version of the first model.
[0016] The PSC may choose any one or more of the above information to instruct the second apparatus to determine the second model, so that the PSC could adapt to different scenarios flexibly.
[0017] In an optional embodiment of the first aspect, the first information includes an ID of the first model, the model information includes an ID of the second model, the determining model information, includes:
[0018] determining the ID of the second model from ID of at least one model.
[0019] The PSC may store the ID of at least one model stored in the second apparatus, wherein the at least one model includes a model that meets requirements of a terminal, i.e., the second model. In this case, the second model could be a partially trained model, it could be provided to the terminal after fine-tuning or a little training, so that the terminal could obtain the second model as soon as possible.
[0020] In an optional embodiment of the first aspect, the at least one model is at least one pre-defined model.
[0021] In the present embodiment, the second model doesn’t need to be trained or fine-tuned, so the second model could be provided to a terminal as soon as possible.
[0022] In an optional embodiment of the first aspect, the method further includes:
[0023] receiving third information from the first apparatus, wherein the third information indicates at least one feature associated with the second model;
[0024] the determining model information, includes:
[0025] determining the model information, according to the third information.
[0026] In this embodiment, the third information may include: preference information (step 1 in Fig. 2) for describing a preferred / desired accuracy and / or a preferred / desired complexity of the second model; and dataset information (step 1 in Fig. 2) for identifying one or more dataset (s) used for training the second model. By the third information, the PSC could determine a terminal’s more accurate requirements for the second model, so that the deployment efficiency and flexibility for the second model could be improved.
[0027] In an optional embodiment of the first aspect, the at least one feature associated with the second model includes an accuracy range and / or a target accuracy of the second model.
[0028] In the present embodiment, the accurate accuracy requirements of a terminal for the second model could be determined, and the accuracy of the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0029] In an optional embodiment of the first aspect, the second model is one whose accuracy is in the accuracy range; and / or, the second model is one whose accuracy is closest to the target accuracy.
[0030] For example, assuming that the accuracy range is represented by [R1, R2] , and three generated candidate models A, B, and C are analyzed to determine the second model. If the accuracy of the model A and the accuracy of the model B are not within the accuracy range [R1, R2] , the accuracy of the model C is within the accuracy range [R1, R2] , the model C is determined as the second model. For another example, assuming that the target accuracy is set as 0.9, the accuracy of the candidate model A is 0.6, the accuracy of the candidate model B is 0.7, and the accuracy of the candidate model C is 0.8, the candidate model C is determined as the second model.
[0031] In an optional embodiment of the first aspect, the at least one feature associated with the second model includes a complexity range and / or a target complexity of the second model.
[0032] In the present embodiment, the accurate complexity requirements of a terminal for the second model could be determined, and the complexity of the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0033] In an optional embodiment of the first aspect, the second model is one whose complexity is in the complexity range; and / or, the second model is one whose complexity is closest to the target complexity.
[0034] For example, assuming that the complexity range is represented by [C1, C2] , three generated candidate models A, B C are analyzed to determine the second model. If the complexity of the model A and the complexity of the model B are not within the complexity range [C1, C2] , the accuracy of the model C is within the complexity range [C1, C2] , the model C is determined as the second model.
[0035] In an optional embodiment of the first aspect, the at least one feature associated with the second model includes at least one model structure, the at least one model structure includes the model structure of the second model.
[0036] In the present embodiment, the accurate model structure requirements of a terminal for the second model could be determined, and the model structure of the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0037] In an optional embodiment of the first aspect, the at least one feature associated with the second model includes information of at least one dataset to be used for training the second model, the at least one dataset includes one or more datasets indicated by the model information.
[0038] In the present embodiment, the accurate training requirements of a terminal for the second model could be determined, and the training of the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0039] In an optional embodiment of the first aspect, the at least one feature associated with the second model includes at least one of the following:
[0040] an ID of an application of the second model, and / or, a scenario ID associated with an application of the second model.
[0041] Different model features are suitable for different applications, by determining the application ID and / or the scenario ID, the features of the second model is more in line with requirements of a terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0042] In an optional embodiment of the first aspect, before the sending second information to the second apparatus, the method further includes:
[0043] sending fourth information to the first apparatus, wherein the fourth information indicates the second model.
[0044] In this embodiment, the fourth information may be, for example, an ID of the second model, that is, a target model ID included in a PSC response (step 3 in Fig. 2) . By the fourth information, a terminal may check whether the second model determined by the PSC is suitable, so that the second model is in line with requirements of the terminal.
[0045] In an optional embodiment of the first aspect, before the sending second information to a second apparatus, the method further includes:
[0046] sending fifth information to the first apparatus, wherein the fifth information indicates accuracy of the second model; and / or, sending sixth information to the first apparatus, wherein the sixth information indicates complexity of the second model.
[0047] In this embodiment, the fifth information may be, for example, model accuracy information included in a PSC response (step 3 in Fig. 2) , and the model accuracy information is used for describing an accuracy of the second model and / or an accuracy difference between the accuracy of the second model and the preferred accuracy of the second model. The sixth information may be, for example, model complexity information included in the PSC response (step 3 in Fig. 2) , and the model complexity information is used for describing a complexity of the second model and / or a complexity difference between the complexity of the second model and the preferred complexity of the second model. By the fifth information, a terminal may check whether the accuracy of the second model is in line with requirements; by the sixth information, a terminal may check whether the complexity of the second model is in line with requirements. So that the deployment efficiency and flexibility for the second model could be improved.
[0048] In an optional embodiment of the first aspect, the fifth information includes a first value, the fifth information indicates accuracy of the second model, includes:
[0049] the first value specifies the percentage or the ratio of number of correct predictions among total number of predictions, wherein the predictions are performed by the second model.
[0050] In an optional embodiment of the first aspect, the sixth information includes a second value and / or a third value, the sixth information indicates complexity of the second model, includes:
[0051] the second value specifies the number of parameters of the second model, and / or, the third value specifies the amount of storage space for storing the second model.
[0052] In an optional embodiment of the first aspect, the method further includes:
[0053] receiving seventh information from the first apparatus, wherein the seventh information indicates the second model is accepted.
[0054] In this embodiment, the seventh information may be, for example, the ID of the second model (i.e., the target model ID) in AF confirmation (step 4 in Fig. 2) . By the seventh information, the PSC could determine the second model is in line with requirements of a terminal.
[0055] In an optional embodiment of the first aspect, the second information is used to configure the second apparatus to perform at least one of the following:
[0056] generating or identifying one or more datasets which are used to train the second model, training the second model according to the first model and the one or more datasets, storing the second model;
[0057] the method further includes:
[0058] receiving eighth information from the second apparatus, the eighth information indicates that the second model has been generated.
[0059] In this embodiment, the eighth information may be, for example, the target model ID included in the MG notification (step 8a in Fig. 2) or in the MH notification (step 8b in Fig. 2) . By the eighth information, the PSC could determine the second model has been generated and then notify a terminal to obtain the second model, so that the second model could be deployed as soon as possible.
[0060] In an optional embodiment of the first aspect, the method further includes:
[0061] receiving ninth information from the second apparatus, the ninth information further indicates an address of the second model, and / or, the ninth information indicates a route to obtain the second model.
[0062] In this embodiment, the ninth information may be, for example, the routing information (step 8 in Fig. 2) . By the ninth information, the PSC could determine the address of the second model and / or the route to obtain the second model, then it could send the address and / or the route to a terminal, so that the terminal could obtain the second model.
[0063] In an optional embodiment of the first aspect, the second information further configures the second apparatus to send the eighth information and the ninth information.
[0064] In the present embodiment, the PSC doesn’t need to configure the second apparatus to send the eighth information and the ninth information through additional messages, thereby saving signaling overhead.
[0065] In an optional embodiment of the first aspect, the method further includes:
[0066] receiving tenth information from the first apparatus, wherein the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to the third apparatus, the third apparatus is used to store the second model.
[0067] In this embodiment, the tenth information may be, for example, the delivery information (step 1 in Fig. 2) or the delivery information (step 104 in Fig. 2) . A terminal may be in a state unable to receive the second model when the second model is generated. For example, the terminal may be powered off or has insufficient storage space when the second model is generated. According to the tenth information, the PSC could notify the second apparatus to promptly send the second model to a temporary storage space (e.g., the third apparatus) , to save storage resources of the second apparatus.
[0068] In an optional embodiment of the first aspect, the method further includes:
[0069] sending eleventh information to the second apparatus, wherein the second apparatus is used to store the second model, the eleventh information includes the tenth information, the eleventh information indicates the second apparatus to send the second model to the third apparatus.
[0070] In this embodiment, the eleventh information may be, for example, the delivery information (step 105 in Fig. 2) . According to the eleventh information, the second apparatus could promptly send the second model to a temporary storage space (e.g., the third apparatus) , to save storage resources.
[0071] In an optional embodiment of the first aspect, the second model is generated by performing Low Rank Adaptation (LoRA) or knowledge distillation.
[0072] In an optional embodiment of the first aspect, the method further includes:
[0073] receiving twelfth information from the first apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .
[0074] In the present embodiment, the accurate requirements of a terminal for the second model could be determined, and the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0075] In an optional embodiment of the first aspect, the twelfth information is provided to the second apparatus when the sending the second information to the second apparatus is performed.
[0076] In an optional embodiment of the first aspect, the performing LoRA is according to the twelfth information.
[0077] In an optional embodiment of the first aspect, the second model is one or more matrices.
[0078] In an optional embodiment of the first aspect, the second apparatus includes a database (DB) , a model generator (MG) , and a model holder (MH) , the DB is used to store at least one dataset to be used for training the second model, the MG is used to generate the second model based on the first model and the at least one dataset, the MH is used to store or hold the second model generated by the MG.
[0079] In an optional embodiment of the first aspect, the DB corresponds to a data collection coordination function (DCCF) or an analytics data repository function (ADRF) , the MG corresponds to a model training logic function (MTLF) , the MH corresponds to the ADRF.
[0080] In an optional embodiment of the first aspect, at least two of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, at least two of the MG, the DB, and the MH correspond to a same entity.
[0081] In an optional embodiment of the first aspect, the method is performed by a fourth apparatus, the fourth apparatus and at least one of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of the MG, the DB, and the MH correspond to a same entity.
[0082] In an optional embodiment of the first aspect, the fourth apparatus is a processing service controller (PSC) , which is used to manage the second apparatus.
[0083] In an optional embodiment of the first aspect, the first apparatus is integrated with a third apparatus, or alternatively, the first apparatus and a third apparatus correspond to a same entity, the third apparatus is used to receive the second model from a MH, and store the second model, the MH is part of the second apparatus.
[0084] In an optional embodiment of the first aspect, the first apparatus is an application function (AF) , the third apparatus is an application server (AS) .
[0085] In a second aspect, an embodiment of the present application provides a method of model processing, the implementer of the method, i.e., a first apparatus, may be a terminal or a server or a chip applied to a server or a terminal, the first apparatus may also be referred to as an application function (AF) . In the following description, a terminal is an example of the first apparatus, not a restriction. The method includes:
[0086] sending first information to a fourth apparatus, wherein the first information indicates a first model;
[0087] receiving fourth information from the fourth apparatus, wherein the fourth information indicates a second model.
[0088] By sending the first information, a PSC could determine the unprocessed model required by a terminal, i.e., the first model, and then determine the processed model corresponding to the first model, i.e., the second model. The PSC could send second information to the second apparatus, to instruct the second apparatus process the first model, the second apparatus may be a model generator (MG) with enough computing and storage resources to process the first model. By the fourth information, the terminal may check whether the second model determined by the PSC is suitable, so that the second model is in line with requirements of the terminal. If the second model is in line with requirements, the terminal may obtain the second model later, so that the terminal could obtain a processed model in case of lacking computing or storage resources.
[0089] In an optional embodiment of the second aspect, the second model is smaller than the first model.
[0090] “The second model is smaller than the first model” may refer to that, the number of parameters of the second model is less than the number of parameters of the first model, and / or, the structure of the second model is simpler than that of the first model. By providing a smaller model to a terminal, the deployment efficiency and flexibility for the second model could be improved.
[0091] In an optional embodiment of the second aspect, before the receiving fourth information from the fourth apparatus, the method further includes:
[0092] sending third information to the fourth apparatus, wherein the third information indicates at least one feature associated with the second model.
[0093] By the third information, a terminal could notify the PSC of its accurate requirements for the second model, so that the deployment efficiency and flexibility for the second model could be improved.
[0094] In an optional embodiment of the second aspect, the at least one feature associated with the second model includes an accuracy range and / or a target accuracy of the second model.
[0095] In the present embodiment, the accurate accuracy requirements of a terminal for the second model could be determined for the PSC, and the accuracy of the second model is more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0096] In an optional embodiment of the second aspect, the second model is one whose accuracy is in the accuracy range; and / or, the second model is one whose accuracy is closest to the target accuracy.
[0097] In an optional embodiment of the second aspect, the at least one feature associated with the second model includes a complexity range and / or a target complexity of the second model.
[0098] In the present embodiment, the accurate complexity requirements of a terminal for the second model could be determined for the PSC, and the complexity of the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0099] In an optional embodiment of the second aspect, the second model is one whose complexity is in the complexity range; and / or, the second model is one whose complexity is closest to the target complexity.
[0100] In an optional embodiment of the second aspect, the at least one feature associated with the second model includes at least one model structure, the at least one model structure includes the model structure of the second model.
[0101] In the present embodiment, the accurate model structure requirements of a terminal for the second model could be determined for the PSC, and the model structure of the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0102] In an optional embodiment of the second aspect, the at least one feature associated with the second model comprises information of at least one dataset to be used for training the second model.
[0103] In the present embodiment, the accurate training requirements of a terminal for the second model could be determined for the PSC, and the training of the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0104] In an optional embodiment of the second aspect, the at least one feature associated with the second model comprises at least one of the following:
[0105] an ID of an application of the second model, and / or, a scenario ID associated with an application of the second model.
[0106] Different model features are suitable for different applications, by determining the application ID and / or the scenario ID, the features of the second model is more in line with requirements of a terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0107] In an optional embodiment of the second aspect, the method further includes:
[0108] receiving fifth information from the fourth apparatus, wherein the fifth information indicates accuracy of the second model; and / or,
[0109] receiving sixth information from the fourth apparatus, wherein the sixth information indicates complexity of the second model.
[0110] By the fifth information, a terminal may check whether the accuracy of the second model is in line with requirements; by the sixth information, a terminal may check whether the complexity of the second model is in line with requirements. So that the deployment efficiency and flexibility for the second model could be improved.
[0111] In an optional embodiment of the second aspect, the fifth information includes a first value, the fifth information indicates accuracy of the second model, includes:
[0112] the first value specifies the percentage or the ratio of number of correct predictions among total number of predictions, wherein the predictions are performed by the second model.
[0113] In an optional embodiment of the second aspect, the sixth information includes a second value and / or a third value, the sixth information indicates complexity of the second model, includes:
[0114] the second value specifies the number of parameters of the second model, and / or, the third value specifies the amount of storage space for storing the second model.
[0115] In an optional embodiment of the second aspect, the method further includes:
[0116] sending seventh information to the fourth apparatus, wherein the seventh information indicates the second model is accepted.
[0117] By the seventh information, the PSC could determine the second model is in line with requirements of a terminal.
[0118] In an optional embodiment of the second aspect, the method further includes:
[0119] sending tenth information to the fourth apparatus, wherein the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to a third apparatus, the third apparatus is used to store the second model.
[0120] A terminal may be in a state unable to receive the second model when the second model is generated. For example, the terminal may be powered off or has insufficient storage space when the second model is generated. According to the tenth information, the PSC could notify the second apparatus to promptly send the second model to a temporary storage space (e.g., the third apparatus) , to save storage resources of the second apparatus.
[0121] In an optional embodiment of the second aspect, the second model is generated by performing LoRA or knowledge distillation.
[0122] In an optional embodiment of the second aspect, the method further includes:
[0123] sending twelfth information to the fourth apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .
[0124] In the present embodiment, the accurate requirements of a terminal for the second model could be determined for the PSC, and the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0125] In an optional embodiment of the second aspect, the performing LoRA is according to the twelfth information.
[0126] In an optional embodiment of the second aspect, the second model is one or more matrices.
[0127] In an optional embodiment of the second aspect, the fourth apparatus and at least one of a model generator (MG) , a database (DB) , and a model holder (MH) are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of a MG, a DB, and a MH correspond to a same entity.
[0128] In an optional embodiment of the second aspect, the fourth apparatus is a processing service controller (PSC) , which is used to manage the MG, the DB, and the MH.
[0129] In an optional embodiment of the second aspect, the method is performed by a first apparatus, the first apparatus is integrated with a third apparatus, or alternatively, the first apparatus and a third apparatus correspond to a same entity, the third apparatus is used to receive the second model from a second apparatus, and store the second model.
[0130] In an optional embodiment of the second aspect, the first apparatus is an application function (AF) , the third apparatus is an application server (AS) .
[0131] In a third aspect, an embodiment of the present application provides a method of model processing, applied to a second apparatus, which may be a server or a chip applied to a server, the second apparatus may also be referred to as a model generator (MG) and / or a model holder (MH) . In the following description, MG or MH is an example of the second apparatus, not a restriction. The method includes:
[0132] receiving second information from a fourth apparatus, wherein the second information is associated with a second model, the second information is used to configure the second apparatus to perform at least one of the following: generating or identifying one or more datasets which are used to train the second model, training the second model according to a first model and the one or more datasets, storing the second model;
[0133] performing steps according to the second information;
[0134] sending eighth information to the fourth apparatus, the eighth information indicates that the second model has been generated.
[0135] By receiving the second information, the second apparatus could determine the processed model what a terminal needs, i.e., the second model. The second apparatus has enough computing and storage resources to obtain the second model, the second model will be sent to the terminal after sending eighth information to the fourth apparatus, so that the terminal could obtain a processed model in case of lacking computing or storage resources.
[0136] In an optional embodiment of the third aspect, the method further includes:
[0137] sending ninth information to the fourth apparatus, the ninth information further indicates an address of the second model, and / or, the ninth information indicates a route to obtain the second model.
[0138] By the ninth information, the PSC could determine the address of the second model and / or the route to obtain the second model, then it could send the address and / or the route to a terminal, so that the terminal could obtain the second model.
[0139] In an optional embodiment of the third aspect, the second information further configures the second apparatus to send the eighth information and the ninth information.
[0140] In the present embodiment, the PSC doesn’t need to configure the second apparatus to send the eighth information and the ninth information through additional messages, thereby saving signaling overhead.
[0141] In an optional embodiment of the third aspect, the second apparatus is used to store the second model, the method further includes:
[0142] receiving eleventh information from the fourth apparatus, wherein the eleventh information comprises tenth information, the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to a third apparatus, the eleventh information indicates the second apparatus to send the second model to the third apparatus, the third apparatus is used to store the second model;
[0143] sending the second model to the third apparatus, according to the eleventh information.
[0144] According to the eleventh information, the second apparatus could promptly send the second model to a temporary storage space (e.g., the third apparatus) , to save storage resources.
[0145] In an optional embodiment of the third aspect, the second information includes at least one of the following:
[0146] an identifier (ID) of the second model, complexity information of the second model, accuracy information of the second information, route information for obtaining the second model, structure information of the second model, information of specifying value of model parameters of the second model, information of one or more datasets to be used for training the second model.
[0147] The PSC may choose any one or more of the above information to instruct the second apparatus to determine the second model, so that the PSC could adapt to different scenarios flexibly.
[0148] In an optional embodiment of the third aspect, the second model is generated by performing LoRA or knowledge distillation.
[0149] In an optional embodiment of the third aspect, the method further includes:
[0150] receiving twelfth information from the fourth apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .
[0151] In the present embodiment, the accurate requirements of a terminal for the second model could be determined, and the second model will be more in line with requirements of the terminal, so that the deployment efficiency and flexibility for the second model could be improved.
[0152] In an optional embodiment of the third aspect, the twelfth information is received from the fourth apparatus when the receiving the second information from the fourth apparatus is performed.
[0153] In an optional embodiment of the third aspect, the performing LoRA is according to the twelfth information.
[0154] In an optional embodiment of the third aspect, the second model is one or more matrices.
[0155] In an optional embodiment of the third aspect, the second apparatus includes a database (DB) , a model generator (MG) , and a model holder (MH) , wherein, the DB is used to store at least one dataset to be used for training the second model, the MG is used to generate the second model based on the first model and the at least one dataset, the MH is used to store or hold the second model generated by the MG.
[0156] In an optional embodiment of the third aspect, the DB corresponds to a data collection coordination function (DCCF) or an analytics data repository function (ADRF) , the MG corresponds to a model training logic function (MTLF) , the MH corresponds to the ADRF.
[0157] In an optional embodiment of the third aspect, at least two of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, at least two of the MG, the DB, and the MH correspond to a same entity.
[0158] In an optional embodiment of the third aspect, the fourth apparatus and at least one of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of the MG, the DB, and the MH correspond to a same entity.
[0159] In an optional embodiment of the third aspect, the fourth apparatus is a processing service controller (PSC) , which is used to manage the second apparatus.
[0160] In a fourth aspect, an embodiment of the present application provides an apparatus of model processing, the apparatus is configured to perform any method provided in the first aspect or the second aspect or the third aspect.
[0161] In a possible design, the apparatus may be divided into functional modules according to any method provided in the first aspect or the second aspect or the third aspect, each functional module may be obtained through division based on a corresponding function, or two or more functions may be integrated into one module.
[0162] For example, the apparatus may be divided into a receiving module, a processing module, a sending module, and the like based on functions. For descriptions of possible technical solutions and advantages performed by the functional modules obtained through division, refer to: the methods provided in the first aspect and the optional implementations thereof, or the methods provided in the second aspect and the optional implementations thereof or the methods provided in the third aspect and the optional implementations thereof.
[0163] In a fifth aspect, an embodiment of the present application provides an apparatus of model processing, the apparatus may be a terminal, a server, or a chip applied to a terminal or a server. The apparatus may comprise a processor, which is used to perform: any method provided in the first aspect and the optional implementations thereof, or any method provided in the second aspect and the optional implementations thereof, or any method provided in the third aspect and the optional implementations thereof.
[0164] Optionally, the apparatus may also comprise a transceiver. When the apparatus is a terminal or a server, the transceiver may be a transceiver circuit, an antenna, etc. When the apparatus is a chip applied to a terminal or a server, the transceiver may be an input / output interface, pin, circuit, etc.
[0165] Optionally, the apparatus may also comprise a memory for storing a computer program or instructions, and the processor executes the computer program or instructions stored in the memory to enable the apparatus to perform: any method provided in the first aspect and the optional implementations thereof, or any method provided in the second aspect and the optional implementations thereof, or any method provided in the third aspect and the optional implementations thereof. When the apparatus is a terminal or a server, the memory may be a read-only memory, random access memory, etc. When the apparatus is a chip applied to a terminal or a server, the memory may be a register, cache, etc.
[0166] In a sixth aspect, an embodiment of the present application provides a system of model processing, the system includes:
[0167] an apparatus provided in the fifth aspect, which is used to perform the method recited in the first aspect; and,
[0168] an apparatus provided in the fifth aspect, which is used to perform the method recited in the second aspect; and,
[0169] an apparatus provided in the fifth aspect, which is used to perform the method recited in the third aspect.
[0170] For example, the system is a communication system, comprising a processing service controller (PSC) , an application function (AF) , and one or multiple network function (s) , wherein, the PSC is used to perform the method recited in the first aspect, the AF is used to perform the method recited in the second aspect, the one or multiple network function (s) are used to perform the method recited in the third aspect.
[0171] For another example, the system is a chip system, comprising at least one processor applied in the above PSC, and at least one processor applied in the above AF, and at least one processor applied in the above one or multiple network function (s) .
[0172] In a seventh aspect, an embodiment of this application provides a computer-readable storage medium, for example, a non-transitory computer-readable storage medium. The computer-readable storage medium stores a computer program or instructions. When the computer program or instructions runs on an apparatus, the apparatus is enabled to perform: any method provided in the first aspect and the optional implementations thereof, or any method provided in the second aspect and the optional implementations thereof, or any method provided in the third aspect and the optional implementations thereof.
[0173] In an eighth aspect, an embodiment of this application provides a computer program product, the computer program product comprises computer program code or computer program instructions, when the computer program code or the computer program instructions executed by an apparatus, the apparatus is enabled to perform: any method provided in the first aspect and the optional implementations thereof, or any method provided in the second aspect and the optional implementations thereof, or any method provided in the third aspect and the optional implementations thereof.
[0174] The apparatus, the system, the computer-readable storage medium, and the computer program product provided in embodiments of this application are all configured to perform the corresponding method provided above. Therefore, regarding the advantages that may be achieved by the apparatus, the system, the computer-readable storage medium, and the computer program product, reference can be made to the advantages of the corresponding method provided above. Details are not repeatedly described again.BRIEF DESCRIPTION OF THE DRAWINGS
[0175] To describe the technical solutions in the embodiments of the present application more clearly, the following briefly describes the accompanying drawings required for describing the embodiments. Apparently, the accompanying drawings in the following description show merely some embodiments of this application, and a person of ordinary skill in the art may derive other drawings from these accompanying drawings without creative efforts.
[0176] FIG. 1 is a system architecture of a system of model processing in accordance with an embodiment of the present application;
[0177] FIG. 2 is a schematic flowchart of a method of model processing implemented based on the system of model processing in FIG. 1 in accordance with an embodiment of the present application;
[0178] FIG. 3 is a schematic flowchart of acquisition of model location information from the processing plane included in the method of FIG. 2 in accordance with an embodiment of the present application;
[0179] FIG. 4 is a schematic flowchart of acquisition of the target model from the MH included in the method of FIG. 2 in accordance with an embodiment of the present application;
[0180] FIG. 5 is a schematic diagram illustrating a more general method of communication in accordance with an embodiment of the present application;
[0181] FIG. 6 is a schematic flowchart illustrating another method of model processing in accordance with an embodiment of the present application;
[0182] FIG. 7 is a schematic block diagram of one apparatus of model processing in accordance with an embodiment of the present application; and
[0183] FIG. 8 is a schematic block diagram of another apparatus of model processing in accordance with an embodiment of the present application.DETAILED DESCRIPTION
[0184] The present application is described in detail below with reference to the accompanying drawings and preferred embodiments. The present application will be further described in detail with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application rather than limiting the present application.
[0185] FIG. 1 illustrates a system architecture of a system of model processing, in order for generating a (likely downsized) target model (i.e., AI model) based on a large pre-trained AI model (referred to as a reference model) according to an embodiment of the present application. The reference model is also referred to as a first model, the target model is also referred to as a second model. The second model may be a variant of the first model or correspond to a version of the first model, and the ID of the second model indicates or identifies the variant of the first model or the version of the first model. As shown in FIG. 1, this system includes a processing service controller (PSC) 101, an application function (AF) 200, a processing plane 200 having one or multiple network function (s) using to process and store models, and an application server (AS) 300. The PSC may be, for example, within a control plane 100. For ease of presentation, it is considered that the DB 201, the MG 203 and the MH 202 constitute a processing plane 200.
[0186] The PSC 101 is configured to obtain information for generating the target model from the AF 400 by performing a PSC-AF negotiation, and the one or multiple network function (s) of the processing plane 200 is / are configured to be managed to perform model processing to enable an apparatus lacking computing or storage resources to obtain a processed model, for example, the one or multiple network function (s) of the processing plane 200 is / are configured to perform model downsizing / compressing to convert the large pre-trained reference model into the downsized target model according to the information for generating the target model obtained from the AF 400.
[0187] In some embodiments, the AF 400 connects to the control plane 100 and interacts / communicates with the PSC 101. The AF 400 may correspond to a wireless terminal device (e.g., a UE) , a network function, or a server. When the AF 400 corresponds to a wireless terminal device, the AF 400 interacts with the PSC 101 via a radio access network (RAN) . The interaction / communication between the AF 400 and the PSC 101 may be implemented through non-access stratum (NAS) messages that are transmitted via the RAN, and the NAS messages may be transmitted and / or received by a control plane function (CPF) in the control plane 100, e.g., the access and mobility management function (AMF) in the 5G system. For example, the AF request (step 1 in Fig. 2) , the AF confirmation (step 4 in Fig. 2) and the AF response (step 104 in Fig. 2) described hereinafter are each included in a NAS message sent from the AF 400 (i.e., the wireless terminal device) via the RAN to the control plane 100. The NAS message may be transported (via the RAN) to the CPF in the control plane 100. In some embodiments, after receiving the NAS message, the CPF forwards the NAS message to the PSC 101, and the PSC 101 receives the NAS message and thus the AF request, the AF confirmation or the AF response in the NAS message. In some embodiments, after receiving the NAS message, the CPF forwards the AF request, the AF confirmation or the AF response in the NAS message sent from the AF 400to the PSC 101, that is, the AF request, the AF confirmation or the AF response in the NAS message (sent from the AF) are received by the CPF, and then is forwarded (via the CPF) to the PSC 101, and the PSC 101 correspondingly receives the AF request, the AF confirmation or the AF response. Alternatively, the NAS message may be transported from the RAN directly to the PSC 101, without involving the CPF as described above, and in this case, the PSC 101 receives the NAS message and thus the AF request, the AF confirmation or the AF response in the NAS message. Likewise, the PSC response (step 3 in Fig. 2) , the PSC notification (step 9 in Fig. 2) described hereinafter are each a NAS message or are each included in a NAS message sent to the AF 400 (i.e., the wireless terminal device) via the RAN. In some embodiments, the PSC 101 sends the PSC response or the PSC notification to the CPF in the control plane 100. After receiving the PSC response or the PSC notification, the CPF generates the NAS message including the PSC response or the PSC notification and sends the NAS message to the AF 400 via the RAN.
[0188] When the AF 400 corresponds to a server or a network function, the AF 400 may interact with the PSC 101 via a network exposure function (NEF) . For example, the AF request (step 1 in Fig. 2) , the AF confirmation (step 4 in Fig. 2) and the AF response (step 104 in Fig. 2) sent from the AF 400 to the PSC 101, the PSC response (step 3 in Fig. 2) and the PSC notification (step 9 in Fig. 2) sent from the PSC 101 to the AF 400, which will be described hereinafter, are each transported between the AF 400 and the PSC 101 via the NEF.
[0189] The PSC 101 is specifically configured to receive an AF request carrying a reference model ID for a target model from the AF and manage the one or multiple network function (s) in the processing plane 200 to perform the model downsizing (the step 6 in FIG. 2) so as to convert the large pre-trained reference model into the downsized target model. The AF request is indicative of a request for generation of the target model. The PSC 101 may be an independent network function.
[0190] The AS 300 is located at a network location and interacts / communicates with the MH 202. The AF 400 and the AS 300 are logical network functions, they may correspond to the same network function or to different network functions.
[0191] The AS 300 may correspond to a server, a wireless terminal device (e.g., a UE) or an independent network function. When the AS 300 interacts / communicates with the MH 202, the interaction / communication between the AS 300 and the MH 202 may be through a data plane path that includes one or multiple DCFs and connects the AS 300 and the MH 202. Data traffic related to the interaction / communication is transported along the data plane path. If the AS 300 corresponds to a wireless terminal device, the interaction / communication between the AS 300 and the MH 202 may further be via the RAN (e.g., through a RAN node that is serving the wireless terminal device) . When the RAN node is serving the wireless terminal device, the RAN node has a wireless connection with the wireless terminal device, and the wireless connection may be implemented through a radio bearer.
[0192] In some embodiments, the PSC 101 is further configured to receive a notification indicating completion of generation of the target model from the processing plane 200 after the target model is generated, and the AS 300 is configured to obtain the target model from the one or multiple network functions (e.g., the MH 202) within the processing plane 200 after the completion of generation of the target model.
[0193] In some embodiments, the PSC 101 is further configured to perform a PSC-AF negotiation between the PSC 101 and the AF 400 to obtain the information for generating the target model. In particular, during a PSC-AF negotiation process, the PSC 101 receives an AF request (e.g., the step 1 in Fig. 2) sent from the AF 400, processes the AF request (e.g., the step 2 in Fig. 2) to obtain a structure of the target model and dataset information (the dataset information is, for example, part of the third information) for generating the target model, and send a PSC response (e.g., the step 3 in Fig. 2) indicating that the AF request has been accepted to the AF 400. The AF request is indicative of a request for generation of the target model.
[0194] In some embodiments, the AF request includes a reference model ID for identifying the reference model, model structure information, and application information for indicating an application or an application scenario that the target model is used for. The reference model ID may comprise information (e.g. a version number) indicating a version of the reference model. The PSC 101 is further configured to determine the structure of the target model according to the model structure information in the AF request. The invention is for a communication system, and the communication system may be, for example, the 5G system. When it is the 5G system, the reference model may be managed by the NWDAF, and information about the reference model (i.e., the reference model ID) is published / provided to the application layer.
[0195] In some embodiments, the PSC 101 is specifically configured to send a PSC configuration request (e.g., the step 5 in Fig. 2) carrying the model structure information and the dataset information to the processing plane 200 in order to manage the one or multiple network function (s) of the processing plane 200 to perform the model downsizing (step 6 in Fig. 2) to convert the large pre-trained reference model into the downsized target model according to the model structure information and the dataset information carried in the PSC configuration request.
[0196] In some embodiments, the application information carried in the AF request indicates an application or an application scenario (or category or class) that the target model is related to or will be used for. For example, the application information includes an application ID or an application scenario (or category or class) ID identifying the application or the application scenario. In some embodiments, the application information further indicates a location (e.g. a geographic area) and / or a time period that the target model is related to or will be used for. For example, the application information includes location information (e.g. one or multiple zone IDs or cell IDs) identifying the location and / or time information (e.g. e.g. time of the day, day of the week, week of the month, and month of the year) identifying the time period. The application information (e.g. the application ID, the application scenario ID, the location information, the time information, or any combination thereof) may correspond to the dataset or correspond to the structure of the target model, which means that the dataset information and the model structure information are optional. If the application information corresponds to the dataset, and if the AF request does not include dataset information, the PSC 101 is configured to process the AF request by identifying the dataset according to the application information that corresponds to the dataset. If the application information corresponds to the structure of the target model, and if the AF request does not include model structure information, the PSC 101 is configured to process the AF request by determining the structure of the target model according to the application information that corresponds to the structure of the target model.
[0197] In some embodiments, the AF request further includes preference information (the preference information is, for example, part of the third information) , and the preference information includes information (referred to as complexity preference information) for describing a desired / preferred complexity of the target model and / or information (referred to as accuracy preference information) describing a desired / preferred accuracy of the target model. The complexity preference information may specify a complexity range, for example, by specifying a maximum number of model parameters that the target model can have or a maximum among storage space that the target model can occupy. A complexity that falls in the complexity range is a desired / preferred complexity. The accuracy preference information may describe a desired / preferred accuracy of the target model by specifying an accuracy range and with respect to or in comparison with the reference model’s accuracy. An accuracy that falls in the accuracy range is a desired / preferred accuracy. In some embodiments, the accuracy information specifies the accuracy range by specifying a percentage or ratio of the accuracy of the reference model as a minimum accuracy that the target model should have. In some embodiments, the accuracy preference information specifies the accuracy range by specifying a percentage or ratio of accuracy that the target model may lose at most (i.e. a maximum accuracy loss or reduction) . The PSC response (e.g., the step 3 in Fig. 2) further includes: information for describing an accuracy of the target model and / or an accuracy difference between the accuracy of the target model and the preferred accuracy of the target model (as described in the preference information) . This information is also referred to as model accuracy information (the model accuracy information, is for example, part of the fifth information) ; and / or information for describing a complexity of the target model and / or a complexity difference between the complexity of the target model and the preferred complexity of the target model (as described in the preference information) . This information is also referred to as the model complexity information (the model complexity information is, for example, part of the sixth information) . In case where only one pre-defined structure corresponds to the application information or is identified in the model structure information, the PSC 101 is configured to determine the structure of the target model by determining the pre-defined structure as the structure of the target model. If there are multiple pre-determined structures corresponding to the application information or identified in the model structure information, the PSC 101 is configured to determine the structure of the target model by: selecting the structure of the target model from the multiple pre-defined structures such that the structure of the target model is associated with an accuracy that is within the accuracy range. Alternatively, the PSC 101 is configured to determine the structure of the target model by selecting the structure of the target model from the multiple pre-defined structures such that the complexity of the target model is within the complexity range.
[0198] In some embodiments, the PSC 101 is further configured to receive an AF confirmation (e.g., the step 4 in Fig. 2) indicating an acceptance of the target model from the AF 400.
[0199] In some embodiments, the PSC 101 is further configured to generate the target model ID (the target model ID is, for example, part of the fourth information) and send the PSC response containing the target model ID to the AF 400 (e.g., step 3 in Fig. 2) if a target model ID for identifying the target model is not provided the AF request. In some embodiments, the AF request may not include the target model ID for identifying the target model.
[0200] In some embodiments, the PSC 101 is further configured to receive the AF confirmation (e.g., the step 4 in Fig. 2) from the AF 400. The AF confirmation may contain delivery information for describing a network address of the AS 300, if the delivery information (the delivery information is, for example, part of the tenth information) is not provided in the AF request. In some embodiments, the delivery information includes routing requirement information associated to a network location of the AS 300. The routing requirement information may include routing information to be used for transporting the target model to a network location, which the AS 300 is located at or corresponds to. The routing information may comprise, for example, a network address, a port number, a protocol name / ID / type, and security credentials associated to the network location. Transporting the target model to the network location causes that the AS 300 receives the target model. The network location of the AS 300 may be identified by an ID (e.g. DNAI in the 5G system) or a network address. In some embodiments, the routing information is pre-configured (e.g., pre-configured in the MH 202) , and the routing requirement information includes a routing profile ID that identifies / maps / corresponds to the routing information. In some embodiments, the delivery information may be provided in the AF request.
[0201] In some embodiments, the one or plurality of network function (s) include (s) a database (DB) 201, a model generator (MG) 203 and a model holder (MH) 202. In some embodiments, the PSC 101 is integrated with one of the network functions (i.e., the DB 201, the MG 203, or the MH 202) as the same network function. In some embodiments, the PSC 101 and the Network functions (i.e., the DB 201, the MG 203, and the MH 202) are integrated as the same network function. In some embodiments, the PSC 101 and any two of the multiple network functions (the DB 201 and the MG 203, or the DB 201 and the MH 202, or the MG 203 and the MH 202) are integrated as the same network function. In some embodiments, the PSC 101 corresponds to the network data analysis function (NWDAF) in the 5G system.
[0202] The DB 201 is configured to store dataset used to generate the target model. In some embodiments, the DB 201 corresponds to the data collection coordination function (DCCF) or the analytics data repository function (ADRF) in the 5G system. Dataset may be managed by the communication system. If the dataset is managed by the system, dataset information may be published / provided to the application layer, and the dataset information may include a dataset ID that maps / corresponds to the dataset and information describing the content of the dataset.
[0203] The MG 203 has access to the reference model and the dataset and is configured to generate the target model based on the reference model. In some embodiments, the MG corresponds to the model training logic function (MTLF) in the 5G system.
[0204] The MH 202 is configured to hold / store the target model after the target model is generated. In some embodiments, the MH 202 corresponds to the analysis data repository function (ADRF) in the 5G system. The DB 201, the MG 203 and the MH 202 are logical network functions and are collectively referred to as network functions or processing service functions (PSFs) .
[0205] In some embodiments, the PSC configuration request includes a first configuration request (e.g., the step 5a in Fig. 2) sent from the PSC 101 to the DB 201, the DB 201 is configured to receive the first configuration request carrying the dataset information, generate or prepare the dataset according to the dataset information and store the generated dataset; in which the dataset information includes a dataset ID / name corresponding to the dataset or comprises information for describing attributes of content of the dataset and information for describing storage location of the content of the dataset, or the dataset information is expressed in a form of one or multiple SQL statement (s) .
[0206] The DB 201 is further configured to send a DB configuration response (e.g. a step subsequent to the step 5a, not shown in Fig. 2) indicating that the configuration request has been received and / or indicating that the dataset has been generated or prepared (ready for use) , in which the DB configuration response includes the dataset ID / name corresponding to the dataset.
[0207] In some embodiments, the PSC configuration request further includes a second configuration request (e.g., the step 5b in Fig. 2) sent from the PSC 101 to the MH 202. The MH 202 is configured to: receive the second configuration request containing a target model ID from the PSC 101 if the AF request does not include delivery information associated to the network location of the AS 300, or receive the second configuration request containing the target model ID and the delivery information from the PSC 101 if the AF request includes the delivery information.
[0208] The MH 202 is further configured to send a MH configuration response (e.g., a step subsequent to the step 5b, not shown in Fig. 2) indicating that the PSC configuration request has been received to the PSC 101. In some embodiments, the MH configuration response includes MH routing information. The MH routing information may include, for example, any of the following: a tunnel ID, a network address (e.g., an IP address, and a port number) , and a protocol name / ID.
[0209] In some embodiments, the PSC configuration request further includes a third configuration request (e.g. the step 5c in Fig. 2) sent from the PSC 101 to the MG 203. The MG 203 is configured to receive the third configuration request carrying the model structure information, the dataset ID / name, and the MH routing information from the PSC 101, generate the target model (the step 6 in Fig. 2) based on the reference model using the model structure information and the dataset identified by the dataset ID / name and obtained from the DB 201, and send the target model and the target model ID to the MH 202 (the step 7 in Fig. 2) according to the MH routing information (the MH routing information is, for example, part of the ninth information) . The MG 203 is further configured to send a MG configuration response (e.g. a step subsequent to the step 5c, not shown in Fig. 2) indicating that the target model has been generated to the PSC 101. In some embodiments, the MG 203 is configured to obtain the dataset from the DB 201 according to the dataset ID / name corresponding to the dataset, and generate the target model (i.e., determine values of model parameters of the target model) using the dataset, for example, in the step 6 of Fig. 2.
[0210] In some embodiment, the MG 203 generates the target model (i.e., determines the values of model parameters of the target model) using the dataset by performing knowledge distillation. During the knowledge distillation, in some embodiments, the reference model is used as a teacher model and the target model is used as a student model, and the student model is trained using the dataset.
[0211] In some embodiments, the MG 203 generates the target model (i.e., determines the values of model parameters of the target model) using the dataset by performing model fine tuning (e.g., Low Rank Adaptation (LoRA) ) . During the model fine turning, the MG 203 fine tunes the reference model using the dataset through LoRA and obtains (generates) a LoRA adaptor as a result. In this case, the LoRA adaptor is the target model. The LoRA adaptor includes one or multiple matrices, and the values of model parameters of the target model are values in the one or multiple matrices. In some embodiments, the MG 203 performs the model fine tuning according to an indication (the indication is, for example, part of the twelfth information received from the PSC 101) . The indication indicates that the target model should be generated through model fine tuning (or in other words, through LoRA) . The PSC 101 may provide the indication to the MG 203 when the PSC 101 configures the MG 203 to generate the target model, for example, in the step 5c in the Fig. 2. The PSC 101 may receive the indication from the AF 400, for example, as part of the AF request in the step 1 in the Fig. 2. In this case, the AF request can be viewed as a request for fine tuning the reference model.
[0212] In some embodiments, the MG 203 is further configured to be notified of a model management event, and the PSC 101 is further configured to: receive a MG notification (the step 8a in the Fig. 2) indicating completion of generation of the target model from the MG 203 when the model management event indicates that the target model has been generated. In some embodiments, the model location information indicating a storage location of the target model and the target model ID (the target model ID is, for example, part of the eighth information) are provided in the MG notification. Alternatively, the PSC 101 is further configured to receive a MH notification (the step 8b in the Fig. 2) indicating the completion of generation of the target model from the MH 202 after the target model is sent from the MG 203 to the MH 202. In some embodiments, the model location information indicating the storage location of the target model and the target model ID (the target model ID is, for example, part of the eighth information) are provided in the MH notification.
[0213] In some embodiments, the MG 203 and the MH 202 correspond to different network functions, the MH 202 is further configured to send acknowledge (ACK) for acknowledging receipt of the target model to the MG 203, the ACK includes model location information indicating the storage location of the target model. The PSC 101 is further configured to: receive the model location information forwarded by the MG 203 (e.g., the MG 203 receives the model location information carried in the ACK from the MH 202 and forwards the model location information to the PSC 101) , or obtain the model location information from the MH configuration response (e.g., the step 5b in Fig. 2) if the MH configuration response includes the model location information, or from the MG notification (e.g., the step 8a in Fig. 2) or from the MH notification (e.g., the step 8b in Fig. 2) ; and send a PSC notification indicating that the target model has been generated to the AF 400 (e.g., the step 9 in Fig. 2) . The PSC notification includes the model location information.
[0214] The AF 400 is further configured to configure / control the AS 300 to obtain the target model from the MH 202 (the step 10 in Fig. 2) after receiving the PSC notification from the PSC 101.
[0215] Various approaches are provided for the AF 400 to manage the AS 300 to obtain the target model from the MH 202.
[0216] In some embodiments, the AF 400 is further configured to send an AF response (e.g., step 104 in Fig. 2) indicating reception of the PSC notification to the PSC 101. The delivery information may be provided in the AF response, and the delivery information includes routing requirement information associated to a network location of the AS 300. The routing requirement information may include routing information to be used for transporting the target model to a network location, which the AS 300 is located at or corresponds to. The routing information may include, for example, a network address, a port number, a protocol name / ID / type, and security credentials associated to the network location. Transporting the target model to the network location causes that the AS 300 receives the target model. The network location of the AS 300 may be identified by an ID (e.g. DNAI in the 5G system) or a network address. In some embodiments, the routing information is pre-configured (e.g., pre-configured in the MH 202) , and the routing requirement information includes a routing profile ID that identifies / maps / corresponds to the routing information. When the AF response include the delivery information (the delivery information is, for example, part of the tenth information) , the PSC 101 is further configured to receive the AF response from the AF 400 and obtain the delivery information carried in the AF response (e.g., step 104 in Fig. 2) and send the delivery information (the delivery information is, for example, part of the eleventh information) to the MH 202 (e.g., step 105 in Fig. 2) to configure / enable the MH 202 to send the target model to the AS 300 (e.g., step 106 in Fig. 2) according to the routing requirement information associated to the network location of the AS 300.
[0217] In some embodiments, the delivery information may be provided in the AF request (step 1 in Fig. 2) or in the AF confirmation (step 4 in Fig. 2) or in the AF response (the step 104 in Fig. 2) , the PSC 101 may be configured to directly obtain the delivery information from the AF request (step 1 in Fig. 2) or from the AF confirmation (e.g., step 103 in Fig. 2) or from the AF response (the step 104 in Fig. 2) , and send the delivery information carried in the AF request (step 1 in Fig. 2) or carried in the AF confirmation (step 4 in Fig. 2) or carried in the AF response (the step 104 in Fig. 2) to the MH 202 (e.g. in step 5b or step 105 in Fig. 2) , in order to manage the MH 202 to send the target model to the AS 300 (e.g., the step 106 in Fig. 2) according to the routing requirement information associated to the network location of the AS 300.
[0218] In some embodiments, the AF 400 is further configured to send the model location information carried in the PSC notification to the AS 300 (e.g., the step 10 in Fig. 2) , in order to manage the AS 300 to download the target model from the MH 202 according to the model location information. The AS 300 may initiate the downloading by sending a model downloading request to the MH 202 according to the model location information.
[0219] In some embodiments, the MH configuration response (e.g., the step 5b in Fig. 2) further includes model location information indicating a storage location of the target model. The PSC 101 is further configured to send a PSC notification (e.g., the step 9 in Fig. 2) indicating that the target model has been generated to the AF 400, and the model location information is provided to the AF 400 in the PSC notification. The AS 300 is configured to be managed to obtain the target model from the MH 202 spontaneously or passively (e.g., the step 10 in Fig. 2) .
[0220] In some embodiments, the PSC 101 is further configured to: receive an AF response (e.g., the step 104 in Fig. 2) indicating reception of the PSC notification from the AF 400, where the AF response includes delivery information, and the delivery information further includes routing requirement information associated to a network location of the AS 300; and send the delivery information to the MH 202 (e.g., the step 105 in Fig. 2) in order that the MH 202 sends the target model to the AS 300 (e.g., the step 106 in Fig. 2) according to the routing requirement information associated to the network location of the AS 300.
[0221] In some embodiments, the routing requirement information may include routing information to be used for transporting the target model to a network location, which the AS 300 is located at or corresponds to. The routing information may comprise, for example, a network address, a port number, a protocol name / ID / type, and security credentials associated to the network location. Transporting the target model to the network location causes that the AS 300 receives the target model. In some embodiments, the routing information is pre-configured (e.g., pre-configured in the MH 202) , and the routing requirement information includes a routing profile ID that identifies / maps / corresponds to the routing information.
[0222] In some embodiments, the AF 400 is further configured to send the model location information carried in the PSC notification (e.g., step 9 in Fig. 2) to the AS 300 so as to manage the AS 300 to download the target model from the MH 202 (e.g., step 101 in Fig. 2) according to the model location information. For example, the AS 300 may send a model downloading request to the MH 202 and downloads the target model from the MH 202 according to the model location information. The AS 300 may download the target model from the MH 202 after the model downloading request is accepted by the MH 202, for example, upon receiving a response from the MH 202 indicating that the model downloading request is accepted.
[0223] In some embodiments, the system further includes a data plane configured to: provide connectivity between any two of the network functions in the processing plane 200, and provide connectivity between the AS 300 and the MH 202 in the processing plane 200. The data plane includes a data connectivity function (DCF) , which is a logical network function and may have one or multiple instances. In some embodiments, the DCF corresponds to the user plane function (UPF) in the 5G system, and the data plane corresponds to the user plane in the 5G system. The DCF provides connectivity among the network functions in the processing plane 200. In other words, interaction / communication between any two of the network functions (e.g., the interaction / communication between the MG 203 and the DB 201, the interaction / communication between the MG 203 and the MH 202) is implemented or supported through the DCF (more precisely, through one or multiple instances of the DCF) . In some embodiments, some (e.g., the MH 202) or all of the network functions are in the data plane. In some embodiments, the MH 202 is in the data plane and is integrated with the DCF (e.g. one of the one or multiple instances of the DCF) of the data plane.
[0224] Corresponding to the system of model processing shown in FIG. 1, FIG. 2 illustrates a schematic flowchart of a method of model processing according to an embodiment of the present application. This method is implemented based on the system of model processing in FIG. 1. More specifically, this method is implemented by the processing service controller (PSC) 101 within the control plane 100 of the system.
[0225] In some embodiments, the method includes:
[0226] receiving an AF request carrying a reference model ID (the reference model ID is, for example, part of the first information) for the target model from the AF 400, which means that, the reference model ID is carried in the AF request, and the reference model ID may be the first information or be part of the first information, in addition to the reference model ID, the AF request may include more information for the target model; and configuring one or multiple network function (s) to generate the target model based on a reference model identified by the reference model ID.
[0227] As shown in FIG. 2, this method involves a PSC-AF negotiation and the PSC-AF negotiation is initiated by the AF 400 and the process of the PSC-AF negotiation includes a step 1, a step 2, a step 3 and a step 4, which are described in detail below:
[0228] At the step 1, the PSC 101 receives the AF request sent from the AF 400, where the AF request is indicative of a request for generation of an AI model (i.e., the target model) based on the reference model (i.e., the large, pre-trained AI model) . The reference model is also referred to as the first model, and the target model is also referred to as the second model. In some embodiments, the AF request may include a target model ID (the target model ID is, for example, part of the third information) for identifying the target model. In some embodiments, the AF request may include delivery information (the delivery information is, for example, part of the tenth information) for describing a network location, which the AS 300 may be located at or corresponds to. In the AF request, the network location may be identified by an ID (e.g. DNAI in the 5G system) or a network address. The delivery information includes routing requirement information associated to the network location. The routing requirement information may include routing information to be used for transporting the target model to the network location. The routing information may comprise, for example, a network address, a port number, a protocol name / ID / type, and security credentials associated to the network location. Transporting the target model to the network location causes that the AS 300 receives the target model. In some embodiments, the routing information is pre-configured (e.g., pre-configured in the MH 202) , and the routing requirement information includes a routing profile ID that identifies / maps / corresponds to the routing information. In some embodiments, the delivery information may be provided in the AF request.
[0229] At the step 2, the PSC 101 processes the AF request to obtain a structure of the target model and dataset information for generating the target model. In some embodiments, the AF request includes the reference model ID for identifying the reference model, model structure information, and dataset information (the dataset information is, for example, part of the third information) for identifying a dataset. Alternatively or additionally, the AF request includes application information for indicating an application or an application scenario that the target model is used for, and the application information includes an application ID or an application scenario ID that corresponds to the dataset or corresponds to the structure of the target model. That is, the model structure information and the dataset information may be optional. The step of processing the AF request in the step 2 includes: determining the structure of the target model according to the model structure information carried in the AF request or according to the application information carried in the AF request.
[0230] The step of processing the AF request in the step 2 includes: identifying the dataset according to the application ID or the application scenario ID that corresponds to the dataset if the AF request does not include the dataset information, and / or determining the structure of the target model according to the application ID or the application scenario ID that corresponds to the structure of the target model if the AF request does not include the model structure information. In some embodiments, the application information further indicates a location (e.g. a geographic area) and / or a time period that the target model is related to or will be used for. For example, the application information includes location information (e.g. one or multiple zone IDs or cell IDs) identifying the location and / or time information (e.g. e.g. time of the day, day of the week, week of the month, and month of the year) identifying the time period. The application information (e.g. the application ID, the application scenario ID, the location information, the time information, or any combination thereof) may correspond to the dataset or correspond to the structure of the target model, and the dataset information and the model structure information may be optional.
[0231] In some embodiments, the AF request further includes preference information (the preference information is, for example, part of the third information) , and the preference information includes information (referred to as complexity preference information) for describing a desired / preferred complexity of the target model in comparison with the reference model’s complexity and / or information (referred to as accuracy preference information) describing a desired / preferred accuracy of the target model in comparison with the reference model’s accuracy. The complexity preference information may specify a complexity range, for example, by specifying a maximum number of model parameters that the target model can have or a maximum among storage space that the target model can occupy. A complexity that falls in the complexity range is a desired / preferred complexity. The accuracy preference information may describe a desired / preferred accuracy of the target model by specifying an accuracy range and with respect to or in comparison with the reference model’s accuracy. An accuracy that falls in the accuracy range is a desired / preferred accuracy. In some embodiments, the accuracy information specifies the accuracy range by specifying a percentage or ratio of the accuracy of the reference model as a minimum accuracy that the target model should have. In some embodiments, the accuracy preference information specifies the accuracy range by specifying a percentage or ratio of accuracy that the target model may lose at most (i.e. a maximum accuracy loss or reduction) . The PSC response (e.g., the step 3 in Fig. 2) further includes: information for describing an accuracy of the target model and / or an accuracy difference between the accuracy of the target model and the preferred accuracy of the target model (as described in the preference information) ; and / or information for describing a complexity of the target model and / or a complexity difference between the complexity of the target model and the preferred complexity of the target model (as described in the preference information) . In case where only one pre-defined structure corresponds to the application information or is identified in the model structure information, the step of processing the AF request in the step 2 includes: determining the pre-defined structure that corresponds to the application information as the structure of the target model. If there are multiple pre-determined structures corresponding to the application information or identified in the model structure information, the step of processing the AF request in the step 2 includes: selecting the structure of the target model from the multiple pre-defined structures such that the structure of the target model is associated with an accuracy that is within the accuracy range; and / or selecting the structure of the target model from the multiple pre-defined structures such that the complexity of the target model is within the complexity range.
[0232] At the step 3, the PSC 101 sends a PSC response indicating that the AF request has been accepted to the AF 400. The PSC response may include: information for describing the accuracy of the target model and / or the accuracy difference between the accuracy of the target model and the preferred accuracy of the target model in the preference information; and / or information for describing the complexity of the target model and / or the complexity difference between the complexity of the target model and the preferred complexity of the target model in the preference information.
[0233] In some embodiments, if the target model ID for identifying the target model is not provided the AF request in the step 1, the step 3 further includes: generating the target model ID and sending the PSC response containing the target model ID to the AF 400.
[0234] At the step 4, the PSC 101 receives an AF confirmation indicating an acceptance of the target model from the AF 400. In some embodiments, the AF confirmation may include the target model ID (the target model ID is, for example, part of the seventh information) . In some embodiments, the AF confirmation may include the delivery information for describing the network location, which the AS 300 is located at or corresponds to. In some embodiments, the AF confirmation may contain the delivery information for describing the network, if the delivery information is not provided in the AF request. In some embodiments, the delivery information includes routing requirement information associated to the network location. The routing requirement information may include routing information to be used for transporting the target model to the network location. The routing information may include, for example, a network address, a port number, a protocol name / ID / type, and security credentials associated to the network location. Transporting the target model to the network location causes that the AS 300 receives the target model. The network location may be identified by an ID (e.g. DNAI in the 5G system) or a network address in the AF request. In some embodiments, the routing information is pre-configured (e.g., pre-configured in the MH 202) , and the routing requirement information includes a routing profile ID that identifies / maps / corresponds to the routing information. The step 4 includes: receiving, if the delivery information for describing the network location is not provided in the AF request, the AF confirmation containing the delivery information from the AF 400.
[0235] In some embodiments, the AF 400 may correspond to a wireless terminal device (e.g., a UE) , a network function, or a server. In the above steps, when the AF 400 corresponds to a wireless terminal device, the AF 400 interacts with the PSC 101 via a radio access network (RAN) . The interaction / communication between the AF 400 and the PSC 101 may be implemented through non-access stratum (NAS) messages that are transmitted via the RAN, and the NAS messages may be transmitted and / or received by a control plane function (CPF) in the control plane 100, e.g., the access and mobility management function (AMF) in the 5G system. For example, the AF request (step 1 in Fig. 2) , the AF confirmation (step 4 in Fig. 2) and the AF response (step 10a in Fig. 2) described hereinafter are each included in a NAS message sent from the AF 400 (i.e., the wireless terminal device) via the RAN to the control plane. The NAS message may be transported (via the RAN) to the CPF in the control plane 100. In some embodiments, after receiving the NAS message the CPF forwards the NAS message to the PSC 101, and the PSC 101 receives the NAS message and thus the AF request, the AF confirmation or the AF response in the NAS message. In some embodiments, after receiving the NAS message the CPF forwards the AF request, the AF confirmation or the AF response in the NAS message (rather than the NAS message) to the NAS message to the PSC 101, and the PSC 101 correspondingly receives the AF request, the AF confirmation or the AF response. Alternatively, the NAS message may be transported from the RAN directly to the PSC 101, without involving the CPF as described above, and in this case, the PSC 101 receives the NAS message and thus the AF request, the AF confirmation or the AF response in the NAS message. Likewise, the PSC response (step 3 in Fig. 2) , the PSC notification (step 9 in Fig. 2) described hereinafter are each a NAS message or are each included in a NAS message sent to the AF 400 (i.e., the wireless terminal device) via the RAN. In some embodiments, the PSC 101 sends the PSC response or the PSC notification to the CPF in the control plane 100. After receiving the PSC response or the PSC notification, the CPF generates the NAS message including the PSC response or the PSC notification and sends the NAS message to the AF 400 via the RAN.
[0236] In some embodiments, when the AF 400 corresponds to a server or a network function, the AF 400 may interact with the PSC 101 via a network exposure function (NEF) . For example, the AF request (step 1 in Fig. 2) , the AF confirmation (step 4 in Fig. 2) and the AF response (step 10a in Fig. 2) sent from the AF 400 to the PSC 101, the PSC response (step 3 in Fig. 2) and the PSC notification (step 9 in Fig. 2) sent from the PSC 101 to the AF 400, which will be described hereinafter, are each transported between the AF 400 and the PSC 101 via the NEF.
[0237] The PSC-AF negotiation, which includes the step 1, the step 2, the step 3 and the step 4, are performed to promote / facilitate the interaction / communication between the PSC 101 and the AF 400, and thereby accelerates the execution speed of the method for generating the downsized target model accordingly.
[0238] In some embodiments, the step of configuring the one or multiple network functions (i.e., the step 5 shown in FIG. 2) includes: sending a PSC configuration request carrying the model structure information and the dataset information to each of the one or multiple network function (s) in the processing plane 200, in order to manage the one or multiple network function (s) to convert (e.g., by downsizing or compressing) the large, pre-trained reference model into the small, customized target model according to the model structure information and the dataset information carried in the PSC configuration request. In some embodiments, the one or multiple network function (s) include (s) a database (DB) 201, a model generator (MG) 203 and a model holder (MH) 202. In some embodiments, the PSC 101 is integrated with one of the network functions (i.e., the DB 201, the MG 203, or the MH202) as the same network function. In some embodiments, the PSC 101 and the Network functions (i.e., the DB 201, the MG 203, and the MH 202) are integrated as the same network function. In some embodiments, the PSC 101 and any two of the network functions (the DB 201 and the MG 203, or the DB 201 and the MH 202, or the MG 203 and the MH 202) are integrated as the same network function. In some embodiments, the PSC 101 corresponds to the network data analysis function (NWDAF) in the 5G system.
[0239] Further referring to FIG. 2, the step of configuring the one or multiple network function (s) further includes a step 5a, a step 5b and a step 5c.
[0240] At the step 5a, the PSC 101 sends a PSC configuration request to a database (DB) 201 in the processing plane 200. The PSC configuration request sent to the DB 201 is also referred to as first configuration request. In some embodiments, the DB 201 corresponds to the data collection coordination function (DCCF) or the analytics data repository function (ADRF) in the 5G system.
[0241] The first configuration request includes the dataset information. According to the dataset information in the first configuration request, the DB 201 may generate or prepare the dataset. In some embodiment, the dataset information includes a dataset ID / name corresponding to the dataset or includes information for describing attributes of content of the dataset and information for describing storage location of the content of the dataset; the dataset information may also be expressed in a form of one or multiple SQL statement (s) ; and
[0242] At the step 5a, the PSC 101 may receive, from the DB 201, a DB configuration response indicating that the first configuration request has been received and / or indicating that the dataset has been generated or prepared (ready for use) . The DB configuration response may include the dataset ID / name corresponding to the dataset.
[0243] At the step 5b, the PSC 101 sends a PSC configuration request to a model holder (MH) 202 in the processing plane 200. The PSC configuration request is also referred to as second configuration request. In some embodiments, the MH 202 corresponds to the analysis data repository function (ADRF) in the 5G system.
[0244] The second configuration request includes the target model ID. The second configuration request may further include the delivery information. In some embodiments, the PSC 101 may configure the MH 202 to perform a step 8b (which will be described in detail hereinafter) by including a notification indication in the second configuration request. If the second configuration request includes the notification indication, the MH 202 will accordingly notify (step 8b in Fig. 2) the PSC 101 about completion of generation of the target model, for example, when or after the MH 202 confirms the completion of generation of the target model.
[0245] At the step 5b, the PSC 101 may receive, from the MH 202, a MH configuration response indicating that the second configuration request has been received. In some embodiments, the MH configuration response includes MH routing information for addressing the MH 202. The MH routing information may include, for example, any of the following: a tunnel ID, a network address (e.g., an IP address and a port number) , and a protocol name / ID. In some embodiments, the MH configuration response may include model location information for describing a storage location of the target model. The model location information may include, for example, any of the following: a domain name (e.g., a fully qualified domain name) , a uniform resource locator (URL) , a network address (e.g., an IP address and a port number) , and a protocol name / ID, such as a hypertext transfer protocol (HTTP) , a file transfer protocol (FTP) .
[0246] At the step 5c, the PSC 101 sends a PSC configuration request to a model generator (MG) 203 in the processing plane 200. The PSC configuration request is also referred to as third configuration request.
[0247] The third configuration request includes the model structure information, the dataset ID / name, and the MH routing information. At the step 5c, the PSC 101 may receive, from the MG 203, a MG configuration response indicating that the third configuration request has been received. Based on the information in the third configuration request, the MG 203 may generate the target model (e.g. by performing knowledge distillation) in the step 6 as further described below. In some embodiments, the PSC 101 may configure the MG 203 to perform a step 8a (which will be described in detail hereinafter) , for example, by including a notification indication in the third configuration request. If the third configuration request includes the notification indication, the MG 203 will accordingly notify (step 8a in Fig. 2) the PSC 101 about completion of generation of the target model, for example, when or after the MG 203 confirms the completion of generation of the target model.
[0248] In some embodiments, this method further includes a step 6 of generating the target model (e.g., via knowledge distillation or Low Rank Adaptation (LoRA) ) .
[0249] At the step 6, the MG 203 converts the reference model into the target model according to the model structure information and the dataset information in the third configuration request. In particular, the MG 203 generates the target model based on the reference model using the model structure information and the dataset identified by the dataset ID / name. The MG 203 obtains the dataset from the DB 201. For example, the MG 203 sends a dataset request to the DB 201, the dataset request including the dataset ID / name. In response to the dataset request, the DB 201 sends the dataset identified by the dataset ID / name to the MG 203. When generating the target model (i.e. converting the reference model to the target model) , the MG 203 determines values of model parameters of the target model by, for example, by performing knowledge distillation or LoRA. During the knowledge distillation, in some embodiments, the reference model is used as a teacher model and the target model is used as a student model, and the student model is trained using the dataset. In some embodiments, the MG 203 generates the target model (i.e., determines the values of model parameters of the target model) using the dataset by performing model fine tuning (e.g., Low Rank Adaptation (LoRA) ) . During the model fine turning, the MG 203 fine tunes the reference model using the dataset through LoRA and obtains (generates) a LoRA adaptor as a result. In this case, the LoRA adaptor is the target model. The LoRA adaptor includes one or multiple matrices, and the values of model parameters of the target model are values in the one or multiple matrices. In some embodiments, the MG 203 performs the model fine tuning according to an indication received from the PSC 101 (the indication is, for example, part of the twelfth information) . The indication indicates that the target model should be generated through model fine tuning (or in other words, through LoRA) . The PSC 101 may provide the indication to the MG 203 when the PSC 101 configures the MG 203 to generate the target model, for example, in the step 5c in the Fig. 2. The PSC 101 may receive the indication from the AF 400, for example, as part of the AF request in the step 1 in the Fig. 2. In this case, the AF request can be viewed as a request for fine tuning the reference model.
[0250] In some embodiments, this method further includes a step 7 of sending the target model and the target model ID to the MH 202.
[0251] At the step 7, the MG 203 sends the target model and the target model ID to the MH 202 according to the MH routing information (the MH routing information is, for example, part of the ninth information) after the generation of the target model.
[0252] Preferably, the method further includes a step 8 of notifying the PSC 101 about the completion of generation of the target model, for example, in a form of model management event (that is, the completion of generation of the target model is a model management event) . The step 8 includes step 8a) notifying, by the MG 203, the PSC 101 about the completion of generation of the target model. In the step 8a) the MG 203 sends an MG notification to the PSC 101, the MG notification indicating the completion of generation of the target model to the PSC 101. The PSC 101 correspondingly receives the MG notification and thus is notified about the completion of generation of the target model. The MG 203 may send the MG notification according to the notification indication (step 5c in Fig. 2) in the third configuration request received from the PSC 101. As an alternative, in some embodiments, the third configuration request (step 5c in Fig. 2) does not include the notification indication, and the MG 203 sends the MG notification to the PSC 101 according to local configuration (which indicates that the MG notification should be sent) . The MG 203 may send the MG notification to the PSC 101 when or after the MG 203 confirms the completion of generation of the target model. The step 8 may further include step 8b) notifying, by the MH 202, the PSC 101 about the completion of generation of the target model. In the step 8b) , the MH 202 sends an MH notification to the PSC 101, the MH notification indicating the completion of generation of the target model. The PSC 101 correspondingly receives the MH notification and is notified about the completion of generation of the target model. The MH 202 may send the MH notification according to the notification indication (e.g., the step 5b in Fig. 2) in the second configuration request received from the PSC 101. In some embodiments, the second configuration request does not include the notification indication, and the MH 202 sends the MH notification to the PSC 101 according to local configuration (which indicates that the MH notification should be sent) . The MH 202 may send the MH notification when or after the MH 202 confirms the completion of generation of the target model. In some embodiment, the MH 202 confirms the completion of generation of the target model when being informed about the completion by the MG 203, for example, in the step 7 in Fig. 2. In some embodiment, the MH 202 confirms the completion when receiving a request for storing the target model or receiving the target model for storage from the MG 203, for example, in the step 7 in Fig. 2. In some embodiments, the step 8 includes only one of the step 8a and the step 8b.
[0253] In some embodiments, the model location information indicating the storage location of the target model and the target model ID (the target model ID is, for example, part of the eighth information) are provided in the MG notification and / or in the MH notification.
[0254] Preferably, the method further includes a step 9 of sending, by the PSC 101, a PSC notification indicating the completion of generation of the target model to the AF 400 (in other words, the target model has been generated) . The AF 400 correspondingly receives the PSC notification (step 9 in Fig. 2) and thus the model location information. In some embodiments, for example, when the MH 202 and the PSC 101 correspond to (or are integrated as) the same network entity, the PSC notification includes the target model (i.e., values of model parameters of the target model) . In this case, when the AF 400 receives the PSC notification from the PSC 101, the AF 400 obtains the target model included in the PSC notification. In some embodiments, the PSC notification may include model location information describing where the target model is stored or how to obtain the target model. In some embodiments, the PSC 101 receives the model location information from the MH 202, for example, as part of the MH configuration response in the step 5b of Fig. 2. As described above, the model location information may include, for example, any of the following: a domain name (e.g., a fully qualified domain name) , a uniform resource locator (URL) , a network address (e.g., an IP address and a port number) , and a protocol name / ID, such as a hypertext transfer protocol (HTTP) , a file transfer protocol (FTP) . The model location information can be used to obtain the target model as further described below, in accordance with the step 10 in the Fig. 2.
[0255] In some embodiment, the PSC response (step 3 in Fig. 2) and the AF confirmation (step 4 in Fig. 2) are optional, and the PSC notification (step 9 in Fig. 2) further includes the information in the PSC response, e.g. the information for describing the accuracy of the target model and / or the accuracy difference between the accuracy of the target model and the preferred accuracy of the target model in the preference information (the preference information is included in the AF request, step 1 in Fig. 2) ; and / or the information for describing the complexity of the target model and / or the complexity difference between the complexity of the target model and the preferred complexity of the target model in the preference information (the preference information is included in the AF request, step 1 in Fig. 2) .
[0256] In some embodiment, the method further includes a step 10 of configuring the AS 300 to obtain the target model from the MH 202.
[0257] In the step 10, the AF 400 provides or sends the model location information to the AS 300. The model location information is included in the PSC notification (step 9 in the Fig. 2) sent from the PSC 101 to the AF 400. In some embodiments, the AS 300 downloads the target model from the MH 202 according to the model location information. For example, the AS 300 may send a model downloading request to the MH 202 and downloads the target model from the MH 202, using the model location information. The AS 300 may download the target model from the MH 202 after the model downloading request is accepted by the MH 202, for example, upon receiving a response from the MH 202 indicating that the model downloading request is accepted. As an alternative, the PSC 101 may receive (the step 104 in the Fig. 2) , from the AF 400, an AF response indicating reception of the PSC notification, wherein the AF response includes delivery information (the delivery information is, for example, part of the tenth information) , and the delivery information further includes routing requirement information associated to a network location, which the AS 300 is located at or corresponds to. In some embodiment, the AF 400 sends the AF response to the PSC 101 in response to the PSC notification received in the step 9 of the Fig. 2. Then, for example, in the step 105 of the Fig. 2, the PSC 101 sends the delivery information carried in the AF response to the MH 202 and configures the MH 202 to send the target model to the network location according to the routing requirement information associated to the network location. As configured by the PSC 101, the MH 202 then, for example, in the step 106 of the Fig. 2, sends the target model according to the routing requirement information such that the AS 300 receives (i.e., obtains) the target model.
[0258] The AS 300 interacts / communicates with the MH 202, for example, to download or receive the target model as described above. The AF 400 and the AS 300 are logical network functions (entities) , they may correspond to the same network function (entity) or to different network functions (entities) . The AS 300 may correspond to a server, a wireless terminal device (e.g., a UE) or an independent network function. When the AS 300 interacts / communicates with the MH 202, the interaction / communication between the AS 300 and the MH 202 may be through a data plane path that includes one or multiple DCFs and connects the AS 300 and the MH 202. Data traffic related to the interaction / communication is transported along the data plane path. If the AS 300 corresponds to a wireless terminal device, the interaction / communication between the AS 300 and the MH 202 may further be via the RAN (e.g., through a RAN node that is serving the wireless terminal device) . When the RAN node is serving the wireless terminal device, the RAN node has a wireless connection with the wireless terminal device, and the wireless connection may be implemented through a radio bearer.
[0259] After the AF 400 or the AS 300 obtains the target model as described above, the AF 400 or the AS 300 can use the target model to perform model inferencing. If the target model is a LoRA adaptor as described above, the LoRA adaptor needs to be loaded into (or combined with) the reference model such that the reference model becomes fine-tuned, and the fine-tuned reference model can then be used to perform the model inferencing. Thus, the AF 400 or the AS 300 requests the system to perform the model inferencing. The request for performing the model inferencing is sent to the PSC 101 and the PSC 101 sends the request to the MG 203. The request for performing the model inferencing may include information identifying the target model (e.g., the target model ID) , and information identifying the reference model (e.g., the reference model ID) . The request may further include the target model (i.e., the LoRA adaptor, which comprises one or multiple matrices) by including the values of model parameters of the target model (i.e. values in the one or multiple matrices) . According to the request for performing the model inferencing, the MG 203 loads the LoRA adaptor into (or combines the LoRA adaptor with) the reference model to make the reference model become fine-tuned. The MG 203 may obtain the LoRA adaptor from within the request (if the request includes the LoRA adaptor) or from the MH 202. The MG 203 then executes the fine-tuned reference model for the model inferencing. The AF 400 or AS 300 can request the system to stop performing the model inferencing (or in other words, to remove the target model from use) . The request for stopping performing the model inferencing is sent to the PSC 101 and the PSC 101 sends the request to the MG 203. The request for stopping performing the model inferencing includes the information identifying the target model and the information identifying the reference model. According to the request for stopping performing the model inferencing, the MG 203 stops executing the fine-tuned reference for the model inferencing, and the MG 203 unloads or separates the LoRA adaptor from the reference model such that the reference model is no longer fine-tuned. As the AF 400 or the AS 300 manages (e.g., stores, transmits) the target model ID and the target model, which is smaller than the reference model, the related management overhead is reduced.
[0260] In some embodiments, when the AF 400 and the AS 300 correspond to the same network function (entity) , the step 10 is an internal process within that network function (entity) . Or, it can be considered that it can be considered the PSC notification in the step 9 is received by the AS 300 and the AS 300 obtains the target model using the model location information in the PSC notification. In this case, the AF response described above can be considered being sent from the AS 300 to the PSC 101 in response to the PSC notification.
[0261] FIG. 3 illustrates a schematic flowchart of acquisition / obtaining of model location information from the one or multiple network function (s) in the processing plane 200 according to an embodiment of the present application. As shown in FIG. 3, after the step 8 of notifying the PSC 101 about the completion of the generation of the target model and before the step 9 of sending the PSC notification to the AF 400, the method further includes a step 891 and a step 892 for obtaining the model location information from the one or multiple network function (s) . The step 891 and the step 892 are executed between the step 8 and the step 9.
[0262] At the step 891, the PSC 101 determines whether model location information indicating the storage location of the target model is included in the MH configuration response (the step 5b in Fig. 2) . If the model location information is included in the MH configuration response (the step 5b in Fig. 2) , the PSC 101 obtains (not shown in Fig. 2) the model location information inform within the MH configuration response (the step 5b in Fig. 2) and thus directly from the MH, and performs the step 9 subsequently. As an alternative, if the model location information is not provided / included in the MH configuration response (the step 5b in Fig. 2) , the step 892 is performed subsequently.
[0263] At the step 892, the PSC 101 obtains the model location information indirectly from the MH 202, via the MG 203. For example, the MG 203 may receive an acknowledge (ACK) for acknowledging receipt of the target model from the MH 202 in the step 7 in Fig. 3. The model location information is provided / included in the ACK, and the MG 203 obtains the model location information from the ACK. Then, the MG 203 sends the model location information to the PSC 101, for example, in the step 8a, and the PSC 101 performs the subsequent step 9.
[0264] In summary, the PSC 101 can obtain the model location information from the MH 202 directly or indirectly in this method of the present application, so that the operating stability of the method is improved. In some embodiments, the MG 203 and the MH 202 may correspond to the same network function (entity) or different network functions (entities) . When the MG 203 and the MH 202 correspond to different network functions (entities) , both the step 891 and the step 892 may need to be performed. When the MG 203 and the MH 202 correspond to the same network function (entity) , the step 891 and the step 892 are internal processes within that network function (entity) .
[0265] Accordingly, FIG. 4 illustrates a schematic flowchart of acquisition of the target model. As shown in FIG. 4, the step 10 of configuring the AS 300 to obtain the target model includes steps 101-107 which are described in detail below:
[0266] In some embodiments, the AS 300 may be managed / enabled / configured to obtain the target model spontaneously via execution of the step 101. At the step 101, the PSC 101 sends the PSC notification to the AF 400, the PSC notification includes the model location information. The step 101 corresponds to the step 9 in Fig. 2. The AF 400 manage / enable / configure the AS 300 to download / obtains the target model in the MH 202according to model location information carried in the PSC notification. As managed / enabled / configured by the AF 400, for example, the AS 300 may send a model downloading request to the MH 202 and download the target model from the MH 202 according to the model location information. The model downloading request may indicate the target model, for example, by including the target model ID. The AS 300 may download (e.g. in the step 106 in Fig. 2) the target model from the MH 202 after the model downloading request is accepted by the MH 202, for example, upon receiving a response from the MH 202 indicating that the model downloading request is accepted.
[0267] In some embodiments, the target model may be obtained by the AS 300 via execution of the steps 102-107. At the step 102, the PSC 101 determines whether the delivery information (the delivery information is, for example, part of the tenth information) is provided in the AF request (the step 1 in Fig. 2) or in the AF confirmation (the step 4 in Fig. 2) . If the delivery information is provided in the AF request (the step 1 in Fig. 2) or provided in the AF confirmation (the step 4 in Fig. 2) , the PSC 101 performs a step 103 subsequently. If the delivery information is neither provided in the AF request (the step 1 of Fig. 2) nor provided in the AF confirmation (the step 4 in Fig. 2) , the PSC 101 performs a step 104 subsequently. In some embodiments, the delivery information includes routing requirement information associated to a network location of the AS 300. The routing requirement information may include routing information to be used for transporting the target model to a network location, which the AS 300 is located at or corresponds to. The routing information may comprise, for example, a network address, a port number, a protocol name / ID / type, and security credentials associated to the network location. Transporting the target model to the network location causes that the AS 300 receives the target model. The network location of the AS 300 may be identified by an ID (e.g. DNAI in the 5G system) or a network address. In some embodiments, the routing information is pre-configured (e.g., pre-configured in the MH 202) , and the routing requirement information includes a routing profile ID that identifies / maps / corresponds to the routing information.
[0268] At the step 103, the PSC 101 obtains the delivery information from / within the AF request (the step 1 in Fig. 2) or from / within the AF confirmation (the step 4 in Fig. 2) directly, if the delivery information is provided / included in the AF request or in the AF confirmation.
[0269] At the step 104, the PSC 101 receives an AF response (the step 104 in Fig. 2) containing the delivery information (the delivery information is, for example, part of the tenth information) from the AF 400. For example, in case where the delivery information is neither provided in the AF request (step 1 in Fig. 2) nor provided in the AF confirmation (step 2 in Fig. 2) , the AF 400 may send the AF response (step 104 in Fig. 2) containing the delivery information to the PSC 101. The AF 400 sends the AF response in response to the PSC notification (step 9 in Fig. 2) received from the PSC 101, acknowledging the receipt of the PSC notification. The PSC 101 correspondingly receives the AF response and obtains the delivery information from / within the AF response.
[0270] At the step 105, the PSC 101 configures the MH 202 to send / transport / deliver the target model to the AS 300 according to the routing requirement information in the delivery information (this delivery information is, for example, part of the eleventh information) . At the step 105, the PSC 101 sends the target model ID and the delivery information (which may be carried / included in the AF request, or in the AF confirmation, or in the AF response, as described above) to the MH 202.
[0271] At the step 106, the MH 202 sends / transports / delivers the target model to the AS 300. The MH 202 may perform this step according to the routing requirement information in the delivery information received in the step 106, or according to the model downloading request received from the AS 300 (as described above) . In the step 106, the MH 202 may further send the target model ID associated to the target model to the AS 300.
[0272] In summary, various approaches are also provided for obtaining the target model in this method. For example, after the PSC notification indicating that the target model has been generated is received by the AF 400, the AS 300 may either be managed / enabled / configured to obtain / download the target model from the one or multiple network function (s) (i.e., the MH 202) spontaneously according to the delivery information received from the AF 400, or passively receive the target model sent from the MH 202. Thus, various approaches for obtaining the target model are provided in this method, and the flexibility and the operating stability of the method is further improved.
[0273] Furthermore, according to the method of the present application, the PSC 101 may obtain the delivery information either from / within the AF request (the step 1 in Fig. 2) or from / within the AF confirmation (the step 4 in Fig. 2) if the delivery information is provided in the AF request or in the AF confirmation, or obtain the delivery information from / within the AF response (the step 104 of the step 10) if the delivery information is provided in the AF response. Thus, various approaches for obtaining the delivery information are also provided in the present application, thus, the approaches for obtaining the target model are further diversified, and the flexibility and the operating stability of the method are further increased.
[0274] Various aspects of the present disclosure are described herein and shown in the drawings by way of example. FIG. 5 is a schematic diagram illustrating a more general method of communication according to embodiments.
[0275] In the method 500, an apparatus may be any device or chip that supports communication and model processing, such as a server or a terminal, or chips applied on a server or a terminal. The forms of the apparatus are not restricted in any embodiment of this application. For ease of presentation, it is considered that the first apparatus is the AF 400, the fourth apparatus is the PSC 101, and the second apparatus is the processing plane 200.
[0276] Different apparatus may be connected through a wireless connection or a wired connection. For example, the first apparatus and the fourth apparatus may be connected through a wireless connection, such as a terrestrial network or a non-terrestrial network. For another example, the fourth apparatus and the second apparatus may be connected through a wired connection, such as a fiber-optic network. It is understood that the apparatus may alternatively be connected through other connection. There is no restriction herein.
[0277] With reference to FIG. 5, the method 500 comprises:
[0278] S510, the first apparatus sends first information to the fourth apparatus, wherein the first information indicates a first model (also referred to as a reference model) .
[0279] Correspondingly, for the fourth apparatus, it receives the first information from the first apparatus.
[0280] The first information may be an ID of the first model, or name of the first model, or any other information which identifies the first model.
[0281] For example, the first information is a 2-bit field in an AF request, when the field is 01, it indicates model 1, when the field is 10, it indicates model 2, wherein model 1 and model 2 are the models that the first apparatus need to use. Correspondingly, the fourth apparatus will determine model 1 as the first model according to the field 01, or determine model 2 as the first model according to the field 10.
[0282] The first model is an unprocessed model required by the first apparatus, for some reasons, such as lacking computing or storage resources, the first apparatus is desirable to convert the first model to a second model to improve deployment efficiency and flexibility.
[0283] For example, the first model may be a large AI model such as LLM or GPT, there is no restriction for the first model. The first apparatus is desirable to downsizing or compressing the large AI model into a small, customized AI model, i.e., the second model.
[0284] For another example, after the system upgrade of the first apparatus, the first model is no longer suitable for the first apparatus, then the first apparatus is desirable to convert the first model into a new model (i.e., the second model) suitable for the new system.
[0285] For another example, the accuracy of the first model is no longer meets customers’ requirements, then the first apparatus is desirable to train the first model into a new model (i.e., the second model) to meet customers’ requirement.
[0286] The second model may be determined by the fourth apparatus, optionally, the first apparatus may also specify the second model by sending information about the second model to the fourth apparatus. This will be discussed in FIG. 6.
[0287] After receiving the first information, the fourth apparatus may perform S520.
[0288] S520, the fourth apparatus determines model information, according to the first information, the model information is associated with a second model (also referred to as a target model) .
[0289] By the first information, the fourth apparatus could determine the unprocessed model required by the first apparatus, and then determine the processed model corresponding to the first model, i.e., the second model.
[0290] In an optional embodiment, the second model is smaller than the first model.
[0291] “The second model is smaller than the first model” may refer to that, the number of parameters of the second model is less than the number of parameters of the first model, and / or, the structure of the second model is simpler than that of the first model. By providing a smaller model to a terminal, the deployment efficiency and flexibility for the second model could be improved.
[0292] In an optional embodiment, the second model is one of at least one model.
[0293] For example, the first information is an ID of the first model, the fourth apparatus stores a plurality of IDs corresponding to a plurality of models, the plurality of models comprise a model that corresponds to the first model, and the model meets requirements of the first apparatus, the fourth apparatus may determine the model as the second model. In this case, the second model could be a partially trained model, it could be provided to the terminal after fine-tuning or a little training.
[0294] In an optional embodiment, the at least one model is at least one pre-defined model.
[0295] In this embodiment, the second model doesn’t need to be trained or fine-tuned, so the second model could be provided to a terminal as soon as possible.
[0296] In an optional embodiment, the second model is a model that need to be trained or fine-tuned based on the first model, in this case, the model information includes at least one of the following:
[0297] an ID of the second model, complexity information of the second model, accuracy information of the second model, route information for obtaining the second model, structure information of the second model, information of specifying value of model parameters of the second model, information of one or more datasets to be used for training the second model.
[0298] The fourth apparatus may choose any one or more of the above information to instruct the second apparatus to determine the second model, i.e., perform the next step.
[0299] S530, the fourth apparatus sends second information to the second apparatus, wherein the second information includes at least part of the model information.
[0300] Correspondingly, for the second apparatus, it receives the second information from the fourth apparatus.
[0301] For the case of pre-defined model (i.e. when the at least one model is at least one pre-defined model) , the second apparatus may determine the second model according to the ID of the second model. The second model may be a variant of the first model or correspond to a version of the first model, and the ID of the second model indicates or identifies the variant of the first model or the version of the first model. As shown in FIG. 1, PSC 101 may send the ID of the second model to the MH 202, which stores a plurality of pre-defined models. MH 202 searches for the second model based on the ID, then MH 202 may send a search result to PSC 101. MH 202 may send the second model to AS 300 according to next instructions from PSC 101, and AF 400 may obtain the second model from AS 300.
[0302] For the case of non pre-defined model, the second apparatus may train or fine-tune the first model to generate the second model. As shown in FIG. 1, PSC 101 may send the model information to MG 203 and DB 201, MG 203 may train or fine-tune the first model based on the model information. After training or fine-tuning, the second model may be stored in MH 202, and MG 203 may send information indicating the second model has been generated to PSC 101, then MH 202 may send the second model to AS 300 according to next instructions from PSC 101, and AF 400 may obtain the second model from AS 300.
[0303] In the two cases of above, it should be noted that AF 400 may also obtain the second model from MH 202 directly.
[0304] In an optional embodiment, for the case of non pre-defined model, the second information is used to configure the second apparatus to perform at least one of the following:
[0305] generating or identifying one or more datasets which are used to train the second model, training the second model according to the first model and the one or more datasets, storing the second model;
[0306] For example, the second model is one or more matrices, such as a convolutional neural network (CNN) , by the second information, the fourth apparatus may configure the second apparatus to identify a picture dataset used to train the CNN.
[0307] For another example, the second model is one or more matrices, such as a generative adversarial network (GAN) , by the second information, the fourth apparatus may configure the second apparatus to generate multiple datasets used to train the GAN.
[0308] After training, the second apparatus may store the second model according to the configuration of the second information. Then the second apparatus may notify the fourth apparatus that the second model has been generated, and / or, the second apparatus may send the second model to an apparatus specified by the fourth apparatus.
[0309] To sum up, the fourth apparatus could send second information to the second apparatus, to instruct the second apparatus process the first model, the second apparatus may be a MG with enough computing and storage resources to process the first model, so that the first apparatus could obtain a processed model in case of lacking computing or storage resources.
[0310] Other optional embodiments of method 500 are introduced below.
[0311] With reference to FIG. 6, the method 500 comprises:
[0312] S511, the first apparatus sends third information to the fourth apparatus, wherein the third information indicates at least one feature associated with the second model.
[0313] Correspondingly, for the fourth apparatus, it receives the third information from the first apparatus, and determines model information according to the third information and the first information, i.e., performs S521.
[0314] It should be noted that, S510, S511, S516, and S517 may be performed simultaneously or asynchronously, which means that, the first information, the third information, the twelfth information, and the tenth information may be in the same message or in different messages.
[0315] For example, the first information and the third information may be in the AF request illustrating in FIG. 2. For another example, the first information, the third information, the twelfth information, and the tenth information may be in the AF request illustrating in FIG. 2.
[0316] When these information are in different messages, there is no restriction on the order in which these information are sent except the first information.
[0317] For example, after the first information be sent, the first apparatus may send the third information first, or send the twelfth information first, or send the tenth information first.
[0318] The third information indicates at least one feature associated with the second model, by the third information, the fourth apparatus could determine more accurate requirements for the second model, so that the deployment efficiency and flexibility for the second model could be improved.
[0319] In an optional embodiment, the at least one feature associated with the second model includes an accuracy range and / or a target accuracy of the second model.
[0320] For example, the accuracy range may be represented by [R1, R2] , wherein R1 and R2 are two endpoints of the accuracy range. For another example, the target accuracy may be one or more values indicating the accuracy of the second model.
[0321] In an optional embodiment, the second model is one whose accuracy is in the accuracy range; and / or, the second model is one whose accuracy is closest to the target accuracy.
[0322] For example, assuming that the accuracy range is represented by [R1, R2] , and three generated candidate models A, B, and C are analyzed to determine the second model. If the accuracy of the model A and the accuracy of the model B are not within the accuracy range [R1, R2] , the accuracy of the model C is within the accuracy range [R1, R2] , the model C is determined as the second model. For another example, assuming that the target accuracy is set as 0.9, the accuracy of the candidate model A is 0.6, the accuracy of the candidate model B is 0.7, and the accuracy of the candidate model C is 0.8, the candidate model C is determined as the second model.
[0323] In the present embodiment, the accuracy of the second model will be more in line with requirements of the first apparatus, so that the deployment efficiency and flexibility for the second model could be improved.
[0324] In an optional embodiment, the at least one feature associated with the second model includes a complexity range and / or a target complexity of the second model.
[0325] The complexity may be the number of layers of the second model, and / or, the complexity may be the number of parameters of the second model. The complexity may also be represented by other parameters, there is no restriction here.
[0326] For example, the complexity range may be represented by [C1, C2] , wherein C1 and C2 are two endpoints of the complexity range. For another example, the target accuracy may be one or more values indicating the complexity of the second model.
[0327] In an optional embodiment, the second model is one whose complexity is in the complexity range; and / or, the second model is one whose complexity is closest to the target complexity.
[0328] For example, assuming that the complexity range is represented by [C1, C2] , three generated candidate models A, B, and C are analyzed to determine the second model. If the complexity of the model A and the complexity of the model B are not within the complexity range [C1, C2] , the accuracy of the model C is within the complexity range [C1, C2] , the model C is determined as the second model.
[0329] In the present embodiment, the accurate complexity requirements for the second model could be determined, and the complexity of the second model will be more in line with requirements of the first apparatus, so that the deployment efficiency and flexibility for the second model could be improved.
[0330] In an optional embodiment, the at least one feature associated with the second model includes at least one model structure, the at least one model structure includes the model structure of the second model.
[0331] The model structure may comprise the relationship between multiple layers, which can be specified by the first apparatus. For example, the model structure may be a convolutional neural network (CNN) , a generative adversarial network (GAN) , a variational auto encoder (VAE) , an echo state network (ESN) , or a long short term memory network (LSTM) .
[0332] In the present embodiment, the accurate model structure requirements for the second model could be determined, and the model structure of the second model will be more in line with requirements of the first apparatus, so that the deployment efficiency and flexibility for the second model could be improved.
[0333] In an optional embodiment, the at least one feature associated with the second model includes information of at least one dataset to be used for training the second model, the at least one dataset includes one or more datasets indicated by the model information.
[0334] The performance of models trained on different dataset will vary. Based on its own requirements, the first apparatus may specify the dataset used to train the second model, so that the training of the second model will be more in line with requirements of the first apparatus.
[0335] For example, when the second model is a CNN, the first apparatus may specify a picture dataset to train the second model. When the second model is a LSTM, the first apparatus may specify a dataset comprising time information to train the second model.
[0336] For another example, when the second model is a GAN, and a GAN for text processing is needed, the first apparatus may specify a picture dataset to train the second model. When the second model is a GAN, and a GAN for video processing is needed, the first apparatus may specify a video dataset to train the second model.
[0337] In an optional embodiment, the at least one feature associated with the second model includes at least one of the following:
[0338] an ID of an application of the second model, and / or, a scenario ID associated with an application of the second model.
[0339] Different model features are suitable for different applications, by determining the application ID and / or the scenario ID, the features of the second model is more in line with requirements of the first apparatus, so that the deployment efficiency and flexibility for the second model could be improved.
[0340] For example, the second model is used to develop application A, the first apparatus may specify the ID of the application A by the third information, so that the features of the second model will be in line with requirements of the application A.
[0341] For another example, the second model is used in an instant messaging scenario, the first apparatus may specify the ID of the instant messaging scenario by the third information, so that the features of the second model will be in line with requirements of the instant messaging scenario.
[0342] In an optional embodiment, before the sending second information to the second apparatus, the method 500 further includes:
[0343] S512, the fourth apparatus sends fourth information to the first apparatus, wherein the fourth information indicates the second model.
[0344] Correspondingly, for the first apparatus, it receives the fourth information from the fourth apparatus, and determines whether the second model could be accepted.
[0345] It should be noted that, S512, S513, and S514 may be performed simultaneously or asynchronously, which means that, the fourth information, the fifth information, and the sixth information may be in the same message or in different messages.
[0346] For example, the fourth information, the fifth information, and the sixth information may be in the PSC response illustrating in FIG. 2.
[0347] When these information are in different messages, there is no restriction on the order in which these information are sent.
[0348] For example, the first apparatus may send the fourth information first, or send the fifth information first, or send the sixth information first.
[0349] The fourth information may be an ID of the second model, or features of the second model, there is no restriction on the form of the fourth information. By the fourth information, the first apparatus may obtain features of the second model, and check whether the second model determined by the fourth apparatus is suitable according to the features of the second model, so that the second model is in line with requirements of the first apparatus.
[0350] In an optional embodiment, before the sending second information to a second apparatus, the method 500 further includes:
[0351] S513, the fourth apparatus sends fifth information to the first apparatus, wherein the fifth information indicates accuracy of the second model.
[0352] S514, the fourth apparatus sends sixth information to the first apparatus, wherein the sixth information indicates complexity of the second model.
[0353] S513 and S514 may both be performed, or only one of the S523 and the S514 is performed.
[0354] Correspondingly, for the first apparatus, it receives the fifth information and / or the sixth information from the fourth apparatus.
[0355] In an optional embodiment, the fifth information includes a first value, the fifth information indicates accuracy of the second model, includes:
[0356] the first value specifies the percentage or the ratio of number of correct predictions among total number of predictions, wherein the predictions are performed by the second model.
[0357] In an optional embodiment, the sixth information includes a second value and / or a third value, the sixth information indicates complexity of the second model, includes:
[0358] the second value specifies the number of parameters of the second model, and / or, the third value specifies the amount of storage space for storing the second model.
[0359] In this embodiment, the fifth information may be, for example, model accuracy information included in a PSC response (step 3 in Fig. 2) , and the model accuracy information is used for describing an accuracy of the second model and / or an accuracy difference between the accuracy of the second model and the preferred accuracy of the second model. The sixth information may be, for example, model complexity information included in the PSC response (step 3 in Fig. 2) , and the model complexity information is used for describing a complexity of the second model and / or a complexity difference between the complexity of the second model and the preferred complexity of the second model.
[0360] By the fifth information, the first apparatus may check whether the accuracy of the second model is in line with requirements; by the sixth information, the first apparatus may check whether the complexity of the second model is in line with requirements. So that the deployment efficiency and flexibility for the second model could be improved.
[0361] In an optional embodiment, the method 500 further includes:
[0362] S515, the first apparatus sends seventh information to the fourth apparatus, wherein the seventh information indicates the second model is accepted.
[0363] Correspondingly, for the fourth apparatus, it receives the seventh information from the first apparatus.
[0364] The seventh information may be a bit, for example, the bit is 1, which indicates the second model is accepted. The seventh information also could be the name of the second model, which indicates the second model is accepted. There is no restriction on the form of the fourth information.
[0365] In this embodiment, the seventh information may be, for example, the ID of the second model (i.e., the target model ID) in AF confirmation (step 4 in Fig. 2) . By the seventh information, the fourth apparatus could determine the second model is in line with requirements of the first apparatus.
[0366] In an optional embodiment, the method 500 further includes:
[0367] S516, the first apparatus sends twelfth information to the fourth apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .
[0368] Correspondingly, for the fourth apparatus, it receives the twelfth information from the first apparatus.
[0369] When the first model needs to be downsized or compressed, there may be multiple ways to achieve this, such as LoRA and knowledge distillation. Different methods are suitable for different scenarios, the first apparatus may specify a method that is suitable for the present scenario by the twelfth information.
[0370] For example, the twelfth information may be an ID of LoRA, or the name of LoRA. There is no restriction on the form of the twelfth information.
[0371] In the present embodiment, the accurate requirements of first apparatus for the second model could be determined, and the second model will be more in line with requirements of the first apparatus, so that the deployment efficiency and flexibility for the second model could be improved.
[0372] In an optional embodiment, the method 500 further includes:
[0373] S531, the fourth apparatus sends the twelfth information to the second apparatus.
[0374] Correspondingly, for the second apparatus, it receives the twelfth information from the fourth apparatus.
[0375] After receiving the twelfth information, the fourth apparatus may send the twelfth information to the second apparatus to indicate the second apparatus to perform LoRA in the process of generating the second model. The twelfth information may be sent with or after the second information.
[0376] After receiving the second information and the twelfth information, the fourth apparatus may perform steps according to the second information and the twelfth information, for example, the fourth apparatus may perform S540 below.
[0377] S540, the second apparatus determine the second model, according to the second information and the twelfth information.
[0378] For example, the second information indicates an ID of a model, the complexity information and accuracy information of the second model, and the twelfth information indicates LoRA, the second apparatus may process the model by LoRA, until the complexity and accuracy of the model meet the requirements of the complexity information and accuracy information, then obtain the second model.
[0379] In an optional embodiment, after obtaining the second model, the second apparatus may perform S532 and / or S533.
[0380] S532, the second apparatus sends eighth information to the fourth apparatus, wherein the eighth information indicates that the second model has been generated.
[0381] Correspondingly, for the fourth apparatus, it receives the eighth information from the second apparatus.
[0382] S533, the second apparatus sends ninth information to the fourth apparatus, the ninth information further indicates an address of the second model, and / or, the ninth information indicates a route to obtain the second model.
[0383] Correspondingly, for the fourth apparatus, it receives the ninth information from the second apparatus.
[0384] The eighth information may be, for example, the target model ID included in the MG notification (step 8a in Fig. 2) or in the MH notification (step 8b in Fig. 2) . Or alternatively, the eighth information may be an acknowledgement (ACK) , there is no restriction on the form of the eighth information.
[0385] By the eighth information, the fourth apparatus could determine the second model has been generated and then notify the first apparatus or the third apparatus to obtain the second model, so that the second model could be deployed as soon as possible.
[0386] The ninth information may be, for example, the routing information (step 8 in Fig. 2) , which indicates a path to MH 202. Or alternatively, the ninth information may be an internet protocol (IP) address of the second apparatus. There is no restriction on the form of the ninth information.
[0387] By the ninth information, the fourth apparatus could determine the address of the second model and / or the route to obtain the second model, then it could send the address and / or the route to the first apparatus or the third apparatus, so that the terminal could obtain the second model.
[0388] In an optional embodiment, the second information further configures the second apparatus to send the eighth information and the ninth information.
[0389] In the present embodiment, the PSC doesn’t need to configure the second apparatus to send the eighth information and the ninth information through additional messages, thereby saving signaling overhead.
[0390] In an optional embodiment, the method 500 further includes:
[0391] S517, the first apparatus sends tenth information to the fourth apparatus, wherein the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to the third apparatus, the third apparatus is used to store the second model.
[0392] Correspondingly, for the fourth apparatus, it receives the tenth information from the first apparatus.
[0393] The tenth information may be, for example, the delivery information (step 1 in Fig. 2) or the delivery information (step 104 in Fig. 2) . The first apparatus may be in a state unable to receive the second model when the second model is generated. For example, the first apparatus may be powered off or has insufficient storage space when the second model is generated. According to the tenth information, the fourth apparatus could notify the second apparatus to promptly send the second model to a temporary storage space (e.g., the third apparatus) , to save storage resources of the second apparatus.
[0394] The address of the third apparatus may be an IP address of the third apparatus, and the route to the third apparatus may be a path from the second apparatus to the third apparatus. There is no restriction on the form of the tenth information.
[0395] In an optional embodiment, after receiving the tenth information, the fourth apparatus may perform S534.
[0396] S534, the fourth apparatus sends eleventh information to the second apparatus, wherein the second apparatus is used to store the second model, the eleventh information includes the tenth information, the eleventh information indicates the second apparatus to send the second model to the third apparatus.
[0397] The eleventh information may be, for example, the delivery information (step 105 in Fig. 2) . According to the eleventh information, the second apparatus could promptly send the second model to a temporary storage space (e.g., the third apparatus) , to save storage resources.
[0398] In an optional embodiment, after obtaining the second model, the second apparatus may perform S550.
[0399] S550, the second apparatus sends the second model to the third model.
[0400] Correspondingly, for the third apparatus, it receives the second model from the second apparatus.
[0401] The second apparatus may send the model to the third apparatus actively, or the second apparatus may send the model to the third apparatus based on a request of the third apparatus. In the former case, the second apparatus should receive the eleventh information; in the latter case, the second apparatus should send the ninth information to the fourth apparatus and the fourth apparatus should forward the ninth information to the third apparatus. There is no restriction on the form of transmission of the second model.
[0402] In an optional embodiment, after obtaining the second model, the third apparatus may perform S560.
[0403] Correspondingly, for the first apparatus, it receives the second model from the third apparatus.
[0404] The third apparatus may send the model to the first apparatus actively, or the third apparatus may send the model to the first apparatus based on a request of the first apparatus. There is no restriction on the form of transmission of the second model.
[0405] For example, the third apparatus determine the first apparatus is powered off or has insufficient storage space at present, the third apparatus may not send the second model actively, but wait for the request of the first apparatus.
[0406] For another example, the third apparatus determine the first apparatus is powered on and has sufficient storage space at present, the third apparatus may not send the second model actively.
[0407] In an optional embodiment, the second apparatus includes a database (DB) , a model generator (MG) , and a model holder (MH) , the DB is used to store at least one dataset to be used for training the second model, the MG is used to generate the second model based on the first model and the at least one dataset, the MH is used to store or hold the second model generated by the MG.
[0408] In an optional embodiment, the DB corresponds to a data collection coordination function (DCCF) or an analytics data repository function (ADRF) , the MG corresponds to a model training logic function (MTLF) , the MH corresponds to the ADRF.
[0409] In an optional embodiment, at least two of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, at least two of the MG, the DB, and the MH correspond to a same entity.
[0410] In an optional embodiment, the method is performed by a fourth apparatus, the fourth apparatus and at least one of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of the MG, the DB, and the MH correspond to a same entity.
[0411] In an optional embodiment, the fourth apparatus is a processing service controller (PSC) , which is used to manage the second apparatus.
[0412] In an optional embodiment, the first apparatus is integrated with a third apparatus, or alternatively, the first apparatus and a third apparatus correspond to a same entity, the third apparatus is used to receive the second model from a MH, and store the second model, the MH is part of the second apparatus.
[0413] In an optional embodiment, the first apparatus is an application function (AF) , the third apparatus is an application server (AS) .
[0414] The examples of methods provided in the embodiments of this application are described in detail in the above, and it is understood that the corresponding devices, in order to achieve the above functions, contain the corresponding hardware structures and / or software modules to implement each function. It should be readily apparent to those skilled in the art that, in combination with the units and algorithmic steps of the examples described in the embodiments disclosed herein, this application may be implemented in hardware or in a combination of hardware and computer software. Whether a function is performed in hardware or computer software drives hardware depends on the specific application and design constraints of the technical solution. Technical professionals may use different methods for each particular application to achieve the described functionality, but such implementation should not be considered beyond the scope of this application.
[0415] FIG. 7 and FIG. 8 are respectively two schematic block diagrams of an apparatus of model processing provided by the embodiments of this application, which may be used to implement the functions of the terminal or server in the embodiments of the above methods, and therefore also have the beneficial effects of the embodiments of the above methods. In embodiments of this application, these apparatuses may be terminals or servers, or modules (e.g., chips) applied to terminals or servers.
[0416] As shown in FIG. 7, apparatus 700 comprises a processing unit 710 and a transceiver unit 720. The transceiver unit 720 performs the receiving step and / or the sending step under the control of the processing unit 710, where the transceiver unit 720 is a sending unit when a sending step is performed, and the transceiver unit 720 is a receiving unit when a receiving step is performed. The apparatus 700 is used to implement the functions of the apparatus in the method embodiment described in the above.
[0417] While the apparatus 700 is used to implement the function of the fourth apparatus in the method 500, the transceiver unit 720 is configured to perform: receiving first information from a first apparatus, wherein the first information indicates a first model; the processing unit 710 is configured to perform: determining model information, according to the first information, the model information is associated with a second model; the transceiver unit 720 is further configured to perform: sending second information to a second apparatus, wherein the second information comprises at least part of the model information.
[0418] Optionally, the second model is smaller than the first model.
[0419] Optionally, the model information comprises at least one of the following: an identifier (ID) of the second model, complexity information of the second model, accuracy information of the second model, route information for obtaining the second model, structure information of the second model, information of specifying value of model parameters of the second model, information of one or more datasets to be used for training the second model. The second model may be a variant of the first model or correspond to a version of the first model, and the ID of the second model indicates or identifies the variant of the first model or the version of the first model.
[0420] Optionally, the first information comprises an ID of the first model, the model information comprises an ID of the second model, the processing unit 710 is further configured to perform: determining the ID of the second model from ID of at least one model.
[0421] Optionally, the at least one model is at least one pre-defined model.
[0422] Optionally, the transceiver unit 720 is further configured to perform: receiving third information from the first apparatus, wherein the third information indicates at least one feature associated with the second model; the processing unit 710 is further configured to perform: determining the model information, according to the third information.
[0423] Optionally, the at least one feature associated with the second model comprises an accuracy range and / or a target accuracy of the second model.
[0424] Optionally, the second model is one whose accuracy is in the accuracy range; and / or, the second model is one whose accuracy is closest to the target accuracy.
[0425] Optionally, the at least one feature associated with the second model comprises a complexity range and / or a target complexity of the second model.
[0426] Optionally, the second model is one whose complexity is in the complexity range; and / or, the second model is one whose complexity is closest to the target complexity.
[0427] Optionally, the at least one feature associated with the second model comprises at least one model structure, the at least one model structure comprises the model structure of the second model.
[0428] Optionally, the at least one feature associated with the second model comprises information of at least one dataset to be used for training the second model, the at least one dataset comprises one or more datasets indicated by the model information.
[0429] Optionally, the at least one feature associated with the second model comprises at least one of the following: an ID of an application of the second model, and / or, a scenario ID associated with an application of the second model.
[0430] Optionally, before the sending second information to the second apparatus, the transceiver unit 720 is further configured to perform: sending fourth information to the first apparatus, wherein the fourth information indicates the second model.
[0431] Optionally, before the sending second information to the second apparatus, the transceiver unit 720 is further configured to perform: sending fifth information to the first apparatus, wherein the fifth information indicates accuracy of the second model; and / or, sending sixth information to the first apparatus, wherein the sixth information indicates complexity of the second model.
[0432] Optionally, the fifth information comprises a first value, the fifth information indicates accuracy of the second model, comprises: the first value specifies the percentage or the ratio of number of correct predictions among total number of predictions, wherein the predictions are performed by the second model.
[0433] Optionally, the sixth information comprises a second value and / or a third value, the sixth information indicates complexity of the second model, comprises: the second value specifies the number of parameters of the second model, and / or, the third value specifies the amount of storage space for storing the second model.
[0434] Optionally, the transceiver unit 720 is further configured to perform: receiving seventh information from the first apparatus, wherein the seventh information indicates that the second model is accepted.
[0435] Optionally, the second information is used to configure the second apparatus to perform at least one of the following: generating or identifying one or more datasets which are used to train the second model, training the second model according to the first model and the one or more datasets, storing the second model; the transceiver unit 720 is further configured to perform: receiving eighth information from the second apparatus, the eighth information indicates that the second model has been generated.
[0436] Optionally, the transceiver unit 720 is further configured to perform: receiving ninth information from the second apparatus, the ninth information further indicates an address of the second model, and / or, the ninth information indicates a route to obtain the second model.
[0437] Optionally, the second information further configures the second apparatus to send the eighth information and the ninth information.
[0438] Optionally, the transceiver unit 720 is further configured to perform: receiving tenth information from the first apparatus, wherein the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to the third apparatus, the third apparatus is used to store the second model.
[0439] Optionally, the transceiver unit 720 is further configured to perform: sending eleventh information to the second apparatus, wherein the second apparatus is used to store the second model, the eleventh information comprises the tenth information, the eleventh information indicates the second apparatus to send the second model to the third apparatus.
[0440] Optionally, the second model is generated by performing Low Rank Adaptation (LoRA) or knowledge distillation.
[0441] Optionally, the transceiver unit 720 is further configured to perform: receiving twelfth information from the first apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .
[0442] Optionally, the twelfth information is provided to the second apparatus when the sending the second information to the second apparatus is performed.
[0443] Optionally, the performing LoRA is according to the twelfth information.
[0444] Optionally, the second model is one or more matrices.
[0445] Optionally, the second apparatus comprises a database (DB) , a model generator (MG) , and a model holder (MH) , the DB is used to store at least one dataset to be used for training the second model, the MG is used to generate the second model based on the first model and the at least one dataset, the MH is used to store or hold the second model generated by the MG.
[0446] Optionally, the DB corresponds to a data collection coordination function (DCCF) or an analytics data repository function (ADRF) , the MG corresponds to a model training logic function (MTLF) , the MH corresponds to the ADRF.
[0447] Optionally, at least two of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, at least two of the MG, the DB, and the MH correspond to a same entity.
[0448] Optionally, the apparatus 700 and at least one of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of the MG, the DB, and the MH correspond to a same entity.
[0449] Optionally, the apparatus 700 is a processing service controller (PSC) , which is used to manage the second apparatus.
[0450] Optionally, the first apparatus is integrated with a third apparatus, or alternatively, the first apparatus and a third apparatus correspond to a same entity, the third apparatus is used to receive the second model from a MH, and store the second model, the MH is part of the second apparatus.
[0451] Optionally, the first apparatus is an application function (AF) , the third apparatus is an application server (AS) .
[0452] While the apparatus 700 is used to implement the function of the first apparatus in the method 500, the transceiver unit 720 is configured to perform: sending first information to a fourth apparatus, wherein the first information indicates a first model; receiving fourth information from the fourth apparatus, wherein the fourth information indicates a second model.
[0453] Optionally, the second model is smaller than the first model.
[0454] Optionally, before the receiving fourth information from the fourth apparatus, the transceiver unit 720 is further configured to perform: sending third information to the fourth apparatus, wherein the third information indicates at least one feature associated with the second model.
[0455] Optionally, the at least one feature associated with the second model comprises an accuracy range and / or a target accuracy of the second model.
[0456] Optionally, the second model is one whose accuracy is in the accuracy range; and / or, the second model is one whose accuracy is closest to the target accuracy.
[0457] Optionally, the at least one feature associated with the second model comprises a complexity range and / or a target complexity of the second model.
[0458] Optionally, the second model is one whose complexity is in the complexity range; and / or, the second model is one whose complexity is closest to the target complexity.
[0459] Optionally, the at least one feature associated with the second model comprises at least one model structure, the at least one model structure comprises the model structure of the second model.
[0460] Optionally, the at least one feature associated with the second model comprises information of at least one dataset to be used for training the second model.
[0461] Optionally, the at least one feature associated with the second model comprises at least one of the following: an ID of an application of the second model, and / or, a scenario ID associated with an application of the second model.
[0462] Optionally, the transceiver unit 720 is further configured to perform: receiving fifth information from the fourth apparatus, wherein the fifth information indicates accuracy of the second model; and / or, receiving sixth information from the fourth apparatus, wherein the sixth information indicates complexity of the second model.
[0463] Optionally, the fifth information comprises a first value, the fifth information indicates accuracy of the second model, comprises: the first value specifies the percentage or the ratio of number of correct predictions among total number of predictions, wherein the predictions are performed by the second model.
[0464] Optionally, the sixth information comprises a second value and / or a third value, the sixth information indicates complexity of the second model, comprises: the second value specifies the number of parameters of the second model, and / or, the third value specifies the amount of storage space for storing the second model.
[0465] Optionally, the transceiver unit 720 is further configured to perform: sending seventh information to the fourth apparatus, wherein the seventh information indicates that the second model is accepted.
[0466] Optionally, the transceiver unit 720 is further configured to perform: sending tenth information to the fourth apparatus, wherein the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to a third apparatus, the third apparatus is used to store the second model.
[0467] Optionally, the second model is generated by performing Low Rank Adaptation (LoRA) or knowledge distillation.
[0468] Optionally, the transceiver unit 720 is further configured to perform: sending twelfth information to the fourth apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .
[0469] Optionally, the performing LoRA is according to the twelfth information.
[0470] Optionally, the second model is one or more matrices.
[0471] Optionally, the fourth apparatus and at least one of a model generator (MG) , a database (DB) , and a model holder (MH) are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of a MG, a DB, and a MH correspond to a same entity.
[0472] Optionally, the fourth apparatus is a processing service controller (PSC) , which is used to manage the MG, the DB, and the MH.
[0473] Optionally, the first apparatus is integrated with a third apparatus, or alternatively, the first apparatus and a third apparatus correspond to a same entity, the third apparatus is used to receive the second model from a second apparatus, and store the second model.
[0474] Optionally, the first apparatus is an application function (AF) , the third apparatus is an application server (AS) .
[0475] While the apparatus 700 is used to implement the function of the second apparatus in the method 500, the transceiver unit 720 is configured to perform: receiving second information from a fourth apparatus, wherein the second information is associated with a second model, the second information is used to configure the second apparatus to perform at least one of the following: generating or identifying one or more datasets which are used to train the second model, training the second model according to a first model and the one or more datasets, storing the second model; the processing unit 710 is configured to perform: steps according to the second information; the transceiver unit 720 is further configured to perform: sending eighth information to the fourth apparatus, wherein the eighth information indicates that the second model has been generated.
[0476] Optionally, the transceiver unit 720 is further configured to perform: sending ninth information to the fourth apparatus, the ninth information further indicates an address of the second model, and / or, the ninth information indicates a route to obtain the second model.
[0477] Optionally, the second information further configures the second apparatus to send the eighth information and the ninth information.
[0478] Optionally, the second apparatus is used to store the second model, the transceiver unit 720 is further configured to perform: receiving eleventh information from the fourth apparatus, wherein the eleventh information comprises tenth information, the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to a third apparatus, the eleventh information indicates the second apparatus to send the second model to the third apparatus, the third apparatus is used to store the second model; sending the second model to the third apparatus, according to the eleventh information.
[0479] Optionally, the second information comprises at least one of the following: an identifier (ID) of the second model, complexity information of the second model, accuracy information of the second information, route information for obtaining the second model, structure information of the second model, information of specifying value of model parameters of the second model, information of one or more datasets to be used for training the second model.
[0480] Optionally, the second model is generated by performing Low Rank Adaptation (LoRA) or knowledge distillation.
[0481] Optionally, the transceiver unit 720 is further configured to perform: receiving twelfth information from the fourth apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .
[0482] Optionally, the twelfth information is received from the fourth apparatus when the receiving the second information from the fourth apparatus is performed.
[0483] Optionally, the performing LoRA is according to the twelfth information.
[0484] Optionally, the second model is one or more matrices.
[0485] Optionally, the second apparatus comprises a database (DB) , a model generator (MG) , and a model holder (MH) , wherein, the DB is used to store at least one dataset to be used for training the second model, the MG is used to generate the second model based on the first model and the at least one dataset, the MH is used to store or hold the second model generated by the MG.
[0486] Optionally, the DB corresponds to a data collection coordination function (DCCF) or an analytics data repository function (ADRF) , the MG corresponds to a model training logic function (MTLF) , the MH corresponds to the ADRF.
[0487] Optionally, at least two of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, at least two of the MG, the DB, and the MH correspond to a same entity.
[0488] Optionally, the fourth apparatus and at least one of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of the MG, the DB, and the MH correspond to a same entity.
[0489] Optionally, the fourth apparatus is a processing service controller (PSC) , which is used to manage the second apparatus.
[0490] The apparatus 700 may be a terminal or a server, or modules (e.g., chips) applied to a terminal or a server. The processing unit 710 may be implemented by hardware or by software. When the processing unit 710 is implemented by hardware, the processing unit 710 is a logic circuit, an integrated circuit, etc. When the processing unit 710 is implemented by software, the processing unit 710 may be a general-purpose processor, implemented by reading software code stored in a memory unit, which may be integrated in the processing unit 710 or may be located outside the processing unit 710 and exist independently.
[0491] As shown in FIG. 8, apparatus 800 includes a processor 810 and an interface circuit 820. The processor 810 and the interface circuit 820 are coupled to each other. It is understood that the interface circuit 820 may be either a transceiver or an input-output interface. Optionally, the apparatus 800 may also include a memory 830 to store the instructions executed by the processor 810, or to store the input data required by the processor 810 to run the instructions, or to store the data produced after the processor 810 has run the instructions.
[0492] When the apparatus 800 is used to implement the method 500 or any other method described in the above, the processor 810 is used to implement the functions of the above processing unit 710 and the interface circuit 820 is used to implement the functions of the above transceiver unit 720.
[0493] When the apparatus 800 is a terminal chip (that is, a chip applied to a terminal) , the terminal chip implements the functions of the terminal in an embodiment of the above method. The terminal chip receives information from an apparatus, which can be understood as the information is first received by other modules in the terminal (such as radio frequency module or antenna) , and then sent to the terminal chip by these modules. The terminal chip sends information to an apparatus, which can be understood as the information is first sent to other modules in the terminal (such as radio frequency module or antenna) , and then sent to the apparatus by these modules.
[0494] When the apparatus 800 is a server chip (i.e., a chip applied to a server) , the server chip implements the functions of the server in an embodiment of the above method. The server chip receives information from an apparatus, which can be understood as the information is first received by other modules in the server (such as the radio frequency module or antenna) , and then sent to the server chip by these modules. The server chip sends information to an apparatus, which can be understood as the information is sent to other modules in the server (such as radio frequency module or antenna) , and then sent to the apparatus by these modules.
[0495] In this application, entity A sends information to entity B, could be from A to B directly or from A to B through other entities. Similarly, entity B receives information from entity A, could be from B to A directly or from B to A through other entities. Here entities A and B may be RAN nodes or terminals, or modules inside the RAN nodes or terminals. The sending and receiving of information may be the information interaction between RAN nodes and terminals, for example, the information interaction between BSs and terminals. The sending and receiving of information can also be the information interaction between two RAN nodes, such as the information interaction between CU and DU. The sending and receiving of information can also be the information interaction between different modules in a device, for example, the information interaction between a terminal chip and other modules in the terminal, or the information interaction between a BS chip and other modules in the BS.
[0496] It is understood that the processor in the embodiments of the present application may be a central processor unit (CPU) , other general-purpose processors, digital signal processors (DSP) , ASIC, FPGA, or any other programmable logic device, transistor logic device, hardware component or any combination thereof. A general-purpose processor may be a microprocessor or any conventional processor.
[0497] The method steps in embodiments of the present application may be implemented in hardware or in software instructions that can be executed by a processor.
[0498] The above embodiments may be implemented in whole or in part by instructions, software, hardware, firmware, or any combination thereof.
[0499] When implemented by instructions, the instructions may be composed of corresponding software modules, the software modules may be stored in a computer readable storage medium. The computer readable storage medium may be a volatile or non-volatile storage medium, or may include both types of volatile and non-volatile storage media. The non-volatile storage medium may be a flash memory, hard disk, mobile hard disk, read-only memory (ROM) , programmable ROM (PROM) , erasable PROM (EPROM) , electrically EPROM (EEPROM) , compact disc ROM (CD-ROM) , or any other form of non-volatile storage medium well known in the art. The volatile storage medium may be a random access memory (RAM) , by illustrative but not restrictive, many forms of RAM are available, such as, registers, static RAM (SRAM) , dynamic RAM (DRAM) , synchronous DRAM (SDRAM) , double data rate SDRAM (DDR SDRAM) , enhanced SDRAM (ESDRAM) , or any other form of volatile storage medium well known in the art.
[0500] An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and can write information to the storage medium. The storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC. Alternatively, the ASIC can be located in a BS or terminal. Processors and storage media can also exist as discrete components in the BS or terminal.
[0501] When implemented by software, 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 programs or instructions. When the computer programs or instructions are loaded and executed on a computer, the process or function described in the embodiment of this application is executed in whole or in part. The computer may be a general computer, a specific computer, a computer network, a network device, an UE or other programmable device. The computer programs or instructions may be stored in or transmitted from one computer readable storage medium to another, for example, the computer programs or instructions may be transmitted from one web site, computer, server or data center to another web site, computer, server or data center by wired or wireless means. The computer readable storage medium may be any available media that the computer can access or a data storage device such as a server or a data center that integrates one or more available media. The available medium may be a magnetic medium, for example, a floppy disk, hard disk, magnetic tape; it can also be an optical medium, for example, a digital video disc; it can also be a semiconductor medium, for example, a solid-state disk (SSD) .
[0502] Finally, regarding the embodiments of this application, there are a few more declarations:
[0503] First, in the embodiments of this application, the first, second and various numerical numbers are only for the purpose of describing convenient distinctions and are not used to limit the scope of the embodiments of this application. For example, the first information and the third information represent two information, which may be two different information or the same information.
[0504] Second, in the embodiments of this application, “indicate” may include direct indicate and indirect indicate, as well as explicit indicate and implicit indicate. The information indicated by a certain information is called the information to be indicated. In the concrete implementation process, there are many ways to indicate the information to be indicated. For example, the information to be indicated can be directly indicated, such as the information to be indicated itself or the index of the information to be indicated. It can also indirectly indicate the information to be indicated by indicating other information, where there is an association relationship between this other information and the information to be indicated. It is also possible to indicate only part of the information to be indicated, and the other parts of the information to be indicated are known or agreed in advance. For example, the indication of the information to be indicated can be realized by pre-agreement (such as protocol stipulation) whether there is an information element, so as to reduce the indication overhead to a certain extent.
[0505] Third, the “protocol” referred to in the embodiments of this application may be a standard protocol in the field of communication, for example, it may include long term evolution (LTE) protocol, new radio (NR) protocol and related protocols in future communication systems, and this application is not limited to this.
[0506] Fourth, “pre-definition” or “pre-configuration” may be achieved by pre-storing the corresponding code, form or other means indicating relevant information in a device (for example, terminal or BS) , and this application does not restrict the specific implementation method. “Store” may mean saved in one or more memories, which may be a separate setting or integrated in the processor or communication device. The one or more memories may also be partially set up separately and partially integrated in the processor or communication device. The type of memories may be any form of storage medium, and this application is not limited to this.
[0507] Fifth, “at least one” means one or more, and “multiple” means two or more. “And / or” refers to the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B, which can represent the situation that A exists alone, B exists alone, and A and B exist simultaneously, where A and B can be a single object or multiple objects. The character “ / ” generally indicates that the associated object is an “or” relationship. “At least one of the following items (items) ” or a similar expression means any combination of these items, including any combination of single or complex items (items) . For example, at least one term (s) of a, b, and c can be expressed as a, or, b, or, c, or, a and b, or, a and c, or, b and c, or, a, b, and c. Where a, b, and c can be a single object or multiple objects, respectively.
[0508] Sixth, in the embodiments of this application, “when ... ” , “at the time of ... ” , “in the case of ... ” , and “if” all refer to an objective circumstance in which a device (for example, terminal or BS) will make the corresponding processing. They do not limit the time, nor do they require the device to have a judgment action in the implementation, nor do they imply other restrictions.
[0509] Seventh, in each embodiment of the application, if there are no special instructions and logical conflicts, the terms and / or descriptions between different embodiments are consistent and can be referred to each other, and the technical features in different embodiments can be combined to form a new embodiment according to their internal logical relationships. The applicant hereby discloses in isolation each individual feature described herein and any combination of two or more such features, to the extent that such features or combinations are capable of being carried out based on the present specification as a whole in the light of the common general knowledge of a person skilled in the art, irrespective of whether such features or combinations of features solve any problems disclosed herein, and without limitation to the scope of the claims. In view of the foregoing description it will be evident to a person skilled in the art that various modifications may be made within the scope of the invention.
[0510] Eighth, although this disclosure refers to illustrative embodiments, this is not intended to be construed in a limiting sense. The description and drawings are, accordingly, to be regarded simply as an illustration of some embodiments of the invention as defined by the appended claims, and are contemplated to cover any and all modifications, variations, combinations or equivalents that fall within the scope of the present invention. Therefore, although embodiments and potential advantages have been described in detail, various changes, substitutions and alterations can be made herein without departing from the invention as defined by the appended claims. Moreover, the scope of the present application is not intended to be limited to the particular embodiments of the process, machine, manufacture, composition of matter, means, methods and steps described in the specification. As one of ordinary skill in the art will readily appreciate from the disclosure of the present invention, processes, machines, manufacture, compositions of matter, means, methods, or steps, presently existing or later to be developed, that perform substantially the same function or achieve substantially the same result as the corresponding embodiments described herein may be utilized according to the present invention. Accordingly, the appended claims are intended to include within their scope such processes, machines, manufacture, compositions of matter, means, methods, or steps.
Claims
1.A method of model processing, comprising:receiving first information from a first apparatus, wherein the first information indicates a first model;determining model information, according to the first information, the model information is associated with a second model;sending second information to a second apparatus, wherein the second information comprises at least part of the model information.2.The method according to claim 1, wherein the second model is smaller than the first model.3.The method according to claim 1 or 2, wherein the model information comprises at least one of the following:an identifier (ID) of the second model, complexity information of the second model, accuracy information of the second model, route information for obtaining the second model, structure information of the second model, information of specifying value of model parameters of the second model, information of one or more datasets to be used for training the second model.4.The method according to any one of claims 1 to 3, wherein the first information comprises an ID of the first model, the model information comprises an ID of the second model, the determining model information, comprises:determining the ID of the second model from ID of at least one model.5.The method according to claim 4, wherein the at least one model is at least one pre-defined model.6.The method according to any one of claims 1 to 5, further comprising:receiving third information from the first apparatus, wherein the third information indicates at least one feature associated with the second model;the determining model information, comprises:determining the model information, according to the third information.7.The method according to claim 6, wherein the at least one feature associated with the second model comprises an accuracy range and / or a target accuracy of the second model.8.The method according to claim 7, wherein:the second model is one whose accuracy is in the accuracy range; and / or,the second model is one whose accuracy is closest to the target accuracy.9.The method according to any one of claims 6 to 8, wherein the at least one feature associated with the second model comprises a complexity range and / or a target complexity of the second model.10.The method according to claim 9, wherein:the second model is one whose complexity is in the complexity range; and / or,the second model is one whose complexity is closest to the target complexity.11.The method according to any one of claims 6 to 10, wherein the at least one feature associated with the second model comprises at least one model structure, the at least one model structure comprises the model structure of the second model.12.The method according to any one of claims 6 to 11, wherein the at least one feature associated with the second model comprises information of at least one dataset to be used for training the second model, the at least one dataset comprises one or more datasets indicated by the model information.13.The method according to any one of claims 6 to 12, wherein the at least one feature associated with the second model comprises at least one of the following:an ID of an application of the second model, and / or, a scenario ID associated with an application of the second model.14.The method according to any one of claims 1 to 13, wherein before the sending second information to the second apparatus, the method further comprises:sending fourth information to the first apparatus, wherein the fourth information indicates the second model.15.The method according to claim 14, wherein before the sending second information to the second apparatus, the method further comprises:sending fifth information to the first apparatus, wherein the fifth information indicates accuracy of the second model; and / or,sending sixth information to the first apparatus, wherein the sixth information indicates complexity of the second model.16.The method according to claim 15, wherein the fifth information comprises a first value, the fifth information indicates accuracy of the second model, comprises:the first value specifies the percentage or the ratio of number of correct predictions among total number of predictions, wherein the predictions are performed by the second model.17.The method according to claim 15, wherein the sixth information comprises a second value and / or a third value, the sixth information indicates complexity of the second model, comprises:the second value specifies the number of parameters of the second model, and / or, the third value specifies the amount of storage space for storing the second model.18.The method according to any one of claims 14-17, further comprising:receiving seventh information from the first apparatus, wherein the seventh information indicates that the second model is accepted.19.The method according to any one of claims 1 to 18, wherein the second information is used to configure the second apparatus to perform at least one of the following:generating or identifying one or more datasets which are used to train the second model, training the second model according to the first model and the one or more datasets, storing the second model;the method further comprises:receiving eighth information from the second apparatus, the eighth information indicates that the second model has been generated.20.The method according to claim 19, wherein the method further comprises:receiving ninth information from the second apparatus, the ninth information further indicates an address of the second model, and / or, the ninth information indicates a route to obtain the second model.21.The method according to claim 20, wherein the second information further configures the second apparatus to send the eighth information and the ninth information.22.The method according to any one of claims 1 to 21, further comprising:receiving tenth information from the first apparatus, wherein the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to the third apparatus, the third apparatus is used to store the second model.23.The method according to claim 22, further comprising:sending eleventh information to the second apparatus, wherein the second apparatus is used to store the second model, the eleventh information comprises the tenth information, the eleventh information indicates the second apparatus to send the second model to the third apparatus.24.The method according to any one of claims 1 to 23, wherein the second model is generated by performing Low Rank Adaptation (LoRA) or knowledge distillation.25.The method according to claim 24, further comprising:receiving twelfth information from the first apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .26.The method according to claim 25, wherein the twelfth information is provided to the second apparatus when the sending the second information to the second apparatus is performed.27.The method according to any one of claims 25-26, wherein the performing LoRA is according to the twelfth information.28.The method according to any one of claims 1 to 27, wherein the second model is one or more matrices.29.The method according to any one of claims 1 to 28, wherein the second apparatus comprises a database (DB) , a model generator (MG) , and a model holder (MH) , the DB is used to store at least one dataset to be used for training the second model, the MG is used to generate the second model based on the first model and the at least one dataset, the MH is used to store or hold the second model generated by the MG.30.The method according to claim 29, wherein the DB corresponds to a data collection coordination function (DCCF) or an analytics data repository function (ADRF) , the MG corresponds to a model training logic function (MTLF) , the MH corresponds to the ADRF.31.The method according to claim 29 or 30, wherein at least two of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, at least two of the MG, the DB, and the MH correspond to a same entity.32.The method according to any one of claims 29 to 31, wherein the method is performed by a fourth apparatus, the fourth apparatus and at least one of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of the MG, the DB, and the MH correspond to a same entity.33.The method according to claim 32, wherein the fourth apparatus is a processing service controller (PSC) , which is used to manage the second apparatus.34.The method according to any one of claims 1 to 33, wherein the first apparatus is integrated with a third apparatus, or alternatively, the first apparatus and a third apparatus correspond to a same entity, the third apparatus is used to receive the second model from a MH, and store the second model, the MH is part of the second apparatus.35.The method according to claim 34, wherein the first apparatus is an application function (AF) , the third apparatus is an application server (AS) .36.A method of model processing, comprising:sending first information to a fourth apparatus, wherein the first information indicates a first model;receiving fourth information from the fourth apparatus, wherein the fourth information indicates a second model.37.The method according to claim 36, wherein the second model is smaller than the first model.38.The method according to claim 36 or 37, wherein before the receiving fourth information from the fourth apparatus, the method further comprises:sending third information to the fourth apparatus, wherein the third information indicates at least one feature associated with the second model.39.The method according to claim 38, wherein the at least one feature associated with the second model comprises an accuracy range and / or a target accuracy of the second model.40.The method according to claim 39, wherein:the second model is one whose accuracy is in the accuracy range; and / or,the second model is one whose accuracy is closest to the target accuracy.41.The method according to any one of claims 38 to 40, wherein the at least one feature associated with the second model comprises a complexity range and / or a target complexity of the second model.42.The method according to claim 41, wherein:the second model is one whose complexity is in the complexity range; and / or,the second model is one whose complexity is closest to the target complexity.43.The method according to any one of claims 38 to 42, wherein the at least one feature associated with the second model comprises at least one model structure, the at least one model structure comprises the model structure of the second model.44.The method according to any one of claims 38 to 43, wherein the at least one feature associated with the second model comprises information of at least one dataset to be used for training the second model.45.The method according to any one of claims 38 to 44, wherein the at least one feature associated with the second model comprises at least one of the following:an ID of an application of the second model, and / or, a scenario ID associated with an application of the second model.46.The method according to any one of claims 36 to 45, further comprising:receiving fifth information from the fourth apparatus, wherein the fifth information indicates accuracy of the second model; and / or,receiving sixth information from the fourth apparatus, wherein the sixth information indicates complexity of the second model.47.The method according to claim 46, wherein the fifth information comprises a first value, the fifth information indicates accuracy of the second model, comprises:the first value specifies the percentage or the ratio of number of correct predictions among total number of predictions, wherein the predictions are performed by the second model.48.The method according to claim 46, wherein the sixth information comprises a second value and / or a third value, the sixth information indicates complexity of the second model, comprises:the second value specifies the number of parameters of the second model, and / or, the third value specifies the amount of storage space for storing the second model.49.The method according to any one of claims 36 to 46, further comprising:sending seventh information to the fourth apparatus, wherein the seventh information indicates that the second model is accepted.50.The method according to any one of claims 36 to 49, further comprising:sending tenth information to the fourth apparatus, wherein the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to a third apparatus, the third apparatus is used to store the second model.51.The method according to any one of claims 36 to 50, wherein the second model is generated by performing Low Rank Adaptation (LoRA) or knowledge distillation.52.The method according to claim 51, further comprising:sending twelfth information to the fourth apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .53.The method according to claim 52, wherein the performing LoRA is according to the twelfth information.54.The method according to any one of claims 37 to 53, wherein the second model is one or more matrices.55.The method according to any one of claims 37 to 54, wherein the fourth apparatus and at least one of a model generator (MG) , a database (DB) , and a model holder (MH) are integrated as one apparatus, or, the fourth apparatus and at least one of a MG, a DB, and a MH correspond to a same entity.56.The method according to claim 55, wherein the fourth apparatus is a processing service controller (PSC) , which is used to manage the MG, the DB, and the MH.57.The method according to any one of claims 37 to 56, wherein the method is performed by a first apparatus, the first apparatus is integrated with a third apparatus, or alternatively, the first apparatus and a third apparatus correspond to a same entity, the third apparatus is used to receive the second model from a second apparatus, and store the second model.58.The method according to claim 57, wherein the first apparatus is an application function (AF) , the third apparatus is an application server (AS) .59.A method of model processing, applied to a second apparatus, the method comprising:receiving second information from a fourth apparatus, wherein the second information is associated with a second model, the second information is used to configure the second apparatus to perform at least one of the following: generating or identifying one or more datasets which are used to train the second model, training the second model according to a first model and the one or more datasets, storing the second model;performing steps according to the second information;sending eighth information to the fourth apparatus, wherein the eighth information indicates that the second model has been generated.60.The method according to claim 59, further comprising:sending ninth information to the fourth apparatus, the ninth information further indicates an address of the second model, and / or, the ninth information indicates a route to obtain the second model.61.The method according to claim 60, wherein the second information further configures the second apparatus to send the eighth information and the ninth information.62.The method according to any one of claims 59 to 61, wherein the second apparatus is used to store the second model, the method further comprises:receiving eleventh information from the fourth apparatus, wherein the eleventh information comprises tenth information, the tenth information indicates an address of a third apparatus, and / or, the tenth information indicates a route to a third apparatus, the eleventh information indicates the second apparatus to send the second model to the third apparatus, the third apparatus is used to store the second model;sending the second model to the third apparatus, according to the eleventh information.63.The method according to any one of claims 59 to 62, wherein the second information comprises at least one of the following:an identifier (ID) of the second model, complexity information of the second model, accuracy information of the second information, route information for obtaining the second model, structure information of the second model, information of specifying value of model parameters of the second model, information of one or more datasets to be used for training the second model.64.The method according to any one of claims 59 to 63, wherein the second model is generated by performing Low Rank Adaptation (LoRA) or knowledge distillation.65.The method according to claim 64, further comprising:receiving twelfth information from the fourth apparatus, wherein the twelfth information indicates that the second model should be generated through model fine tuning (or in other words, through LoRA) .66.The method according to claim 65, wherein the twelfth information is received from the fourth apparatus when the receiving the second information from the fourth apparatus is performed.67.The method according to any one of claims 65-66, wherein the performing LoRA is according to the twelfth information.68.The method according to any one of claims 60 to 67, wherein the second model is one or more matrices.69.The method according to any one of claims 60 to 68, wherein the second apparatus comprises a database (DB) , a model generator (MG) , and a model holder (MH) , wherein, the DB is used to store at least one dataset to be used for training the second model, the MG is used to generate the second model based on the first model and the at least one dataset, the MH is used to store or hold the second model generated by the MG.70.The method according to claim 69, wherein the DB corresponds to a data collection coordination function (DCCF) or an analytics data repository function (ADRF) , the MG corresponds to a model training logic function (MTLF) , the MH corresponds to the ADRF.71.The method according to claim 69 or 70, wherein at least two of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, at least two of the MG, the DB, and the MH correspond to a same entity.72.The method according to any one of claims 69 to 71, wherein the fourth apparatus and at least one of the MG, the DB, and the MH are integrated as one apparatus, or alternatively, the fourth apparatus and at least one of the MG, the DB, and the MH correspond to a same entity.73.The method according to claim 72, wherein the fourth apparatus is a processing service controller (PSC) , which is used to manage the second apparatus.74.An apparatus of model processing, the apparatus comprises:at least one processor; andat least one memory, coupled to the at least one processor and configured to store instructions that, when executed by the at least one processor, cause the apparatus to:perform the method according to any of claims 1 to 73.75.A system of model processing, the system comprises:an apparatus provided in claim 71 which is used to perform the method according to any of claims 1 to 35; and,an apparatus provided in claim 73, which is used to perform the method according to any of claims 36 to 58; and,an apparatus provided in claim 73, which is used to perform the method according to any of claims 59 to 73.
Citation Information
Patent Citations
Neural network compression method, image processing method and device
CN110717578A
Customized deep neural network model compression method and system based on cloud edge cooperation
CN112486686A
Edge calculation method and system based on model compression and service distribution
CN112860411A
Model transmission method, terminal and network side equipment
CN116827497A
Apparatuses and methods for sharing trained machine learning models
EP4239522A1