Detailed Description
The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.
The method for processing model tasks provided by the embodiment of the application can be applied to an application environment shown in fig. 1, wherein the application environment can comprise a terminal, an intelligent wireless network (AI RAN, artificial Intelligence Radio Access Network) node and a wireless intelligent management and arrangement function (wireless intelligent management and arrangement function layer, RAN AI LAYER). The terminal can be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, internet of things equipment and portable wearable equipment, and the internet of things equipment can be smart speakers, smart televisions, smart air conditioners, smart vehicle-mounted equipment, projection equipment and the like. The portable wearable device may be a headset device or the like. The head-mounted device may be a Virtual Reality (VR) device, an augmented Reality (Augmented Reality, AR) device, or the like.
The intelligent wireless network can apply artificial intelligence and machine learning technology to various links such as planning, deployment, management and optimization of a wireless access network, so as to improve network performance and resource utilization efficiency, enhance flexibility and intelligence of the network, and better meet requirements of users on high-speed, stable and low-delay network connection.
The wireless intelligent management and arrangement function can conduct intelligent and automatic management and scheduling on the wireless network, network performance is improved, resource utilization is optimized, service flexibility is enhanced, and user experience is improved. As shown in fig. 2, the wireless intelligent management and orchestration function may comprise a general computing resource scheduling module, which may be a component of the wireless intelligent management and orchestration function that implements efficient collaborative management and optimal allocation of resources such as computation, communication, etc. in a wireless communication network, and may be connected to various intelligent wireless networks, including, but not limited to, a 3GPP RAN (a radio access network in the third generation partnership project 3 GPP), a Yun Yuansheng RAN (a network architecture that combines cloud native technology with the radio access network RAN).
In the application environment shown in fig. 1, a terminal can access a node of an intelligent wireless network (called an intelligent wireless network node), and the intelligent wireless network node accessed by the terminal is denoted as an access node. The intelligent radio network may comprise a plurality of intelligent radio network nodes (intelligent radio network nodes 1 to n) which may be controlled and managed by a radio intelligent management and orchestration function.
In the related art, when facing performance bottlenecks, a terminal generally offloads a model reasoning task to an edge computing (MEC, mobile Edge Computing) node for execution, and migrates the model reasoning task of a deep learning model to an edge node with stronger computation power for execution, but the computation power of the edge computing is deployed at the edge node of a network, the model task processing needs to be completed through network forwarding, and a large communication delay is introduced in data transmission between the terminal and an external node, so that the technical problem of low timeliness of the model task processing exists, and the task requirement of low delay cannot be met.
In this regard, the method for processing model tasks provided by the embodiment of the application can divide the model reasoning task part of the terminal into the intelligent wireless network nodes, the terminal and the intelligent wireless network nodes perform collaborative reasoning to complete the model reasoning task, and the intelligent wireless network nodes perform collaborative reasoning to reduce the transmission paths of related data in architecture, thereby reducing the communication time delay and improving the timeliness of model task processing.
The model task processing method of the present application will be described below with reference to various embodiments and corresponding drawings based on the application environment as shown in fig. 1.
In an exemplary embodiment, as shown in fig. 3, a model task processing method is provided, which may be performed by the wireless intelligent management and orchestration function shown in fig. 1, and may include the steps of:
step S301, a model joint reasoning request sent by an access node is obtained.
In this step, the wireless intelligent management and orchestration function may obtain a model joint reasoning request sent by the access node.
The access node refers to an intelligent wireless network node to which the terminal is accessed, as shown in fig. 1, the access node may be an intelligent wireless network node 1, the access node is connected with a wireless intelligent management and arrangement function, the wireless intelligent management and arrangement function may control and manage the access node, and the wireless intelligent management and arrangement function may receive a model joint reasoning request from the access node.
The model joint reasoning request is sent to the access node by the terminal, and the access node can forward the model joint reasoning request to the wireless intelligent management and arrangement function. The model joint reasoning request is used for requesting the terminal to assist the model reasoning from the intelligent wireless network node to the wireless intelligent management and arrangement function, and can also be called as assisting in processing the model reasoning task. The model can comprise a machine learning model, a deep learning model and the like, can be determined according to a specific model reasoning task which is requested to assist by the terminal, can indicate a specific model in a model joint reasoning request, can also be determined for the terminal by a wireless intelligent management and arrangement function and the like.
The model joint reasoning request can contain joint reasoning requirements, and the terminal can package the joint reasoning requirements to obtain the model joint reasoning request. The joint reasoning requirement can be the requirement of the model reasoning task, and can comprise the requirements of calculation power, time delay and the like.
Step S302, determining an inference node meeting the joint inference requirement in the intelligent wireless network node set.
The wireless intelligent management and arrangement function can determine the intelligent wireless network node set according to the coverage area/management range of the wireless intelligent management and arrangement function to the intelligent wireless network node. The intelligent radio network node set may comprise intelligent radio network nodes within the coverage/management range of the wireless intelligent management and orchestration function, the intelligent radio network node set comprising the above-mentioned access nodes. In this step, the wireless intelligent management and orchestration function may determine, from the set of intelligent wireless network nodes, one or more intelligent wireless network nodes that meet the above-described joint reasoning requirement as reasoning nodes. The inference node is one or more intelligent wireless network nodes in the intelligent wireless network node set, which are used for assisting the terminal in carrying out model inference. As described above, the joint reasoning requirements may include requirements in terms of computational effort, time delay, etc., and the reasoning nodes may be one or more intelligent wireless network nodes that meet the joint reasoning requirements in terms of computational effort, time delay, etc.
Step S303, triggering the reasoning node to execute the model joint reasoning task corresponding to the model joint reasoning request and transmitting the node reasoning result to the terminal through the access node so as to enable the terminal to obtain the model reasoning result.
In this step, the wireless intelligent management and orchestration function may send a model joint reasoning task corresponding to the model joint reasoning request to the reasoning node, trigger the reasoning node to execute the model joint reasoning task, and transmit a node reasoning result obtained by executing the model joint reasoning task to the terminal through the access node, so that the terminal obtains the model reasoning result. The model reasoning result is obtained based on the terminal reasoning result and the node reasoning result of the terminal.
The wireless intelligent management and arrangement function can determine a corresponding model joint reasoning task according to the model joint reasoning request, the model joint reasoning task refers to a model reasoning task which needs the reasoning node to assist in execution, the task belongs to a part of the model reasoning task of the terminal, the model joint reasoning task can be contained in the model joint reasoning request and can be obtained by dividing the model joint reasoning request by the terminal, and the wireless intelligent management and arrangement function can also be obtained by dividing the model joint reasoning task of the terminal according to joint reasoning requirements in the model joint reasoning request.
The number of the inference nodes can be multiple, each inference node can execute respective model joint inference tasks, and each inference node can transmit node inference results obtained after executing respective model joint inference tasks to the terminal through the access node.
Therefore, the terminal can obtain the reasoning node through the access node, so that the model reasoning result of the model reasoning task of the terminal can be obtained based on the terminal reasoning result and the node reasoning result of the terminal. The terminal reasoning result refers to a reasoning result obtained by the terminal executing a model local reasoning task, the model local reasoning task and the model joint reasoning task can form a model reasoning task of the terminal, and the model local reasoning task can be obtained by the terminal or can be obtained by the wireless intelligent management and arrangement function in the model reasoning task of the terminal.
According to the model task processing method, a wireless intelligent management and arrangement function can acquire a model joint reasoning request sent by an access node, the access node is an intelligent wireless network node accessed by a terminal, the model joint reasoning request is sent to the access node by the terminal, the model joint reasoning request comprises joint reasoning requirements, the wireless intelligent management and arrangement function determines a reasoning node meeting the joint reasoning requirements in an intelligent wireless network node set, the intelligent wireless network node set comprises the access node, the wireless intelligent management and arrangement function can trigger the reasoning node to execute a model joint reasoning task corresponding to the model joint reasoning request and transmit a node reasoning result to the terminal through the access node so as to obtain the model reasoning result by the terminal, and the model reasoning result is obtained based on the terminal reasoning result and the node reasoning result of the terminal. According to the scheme, the model reasoning task of the terminal can be divided into the intelligent wireless network nodes, the model reasoning task is completed through the cooperative reasoning of the terminal and the intelligent wireless network nodes, and the transmission path of related data can be reduced in a framework through the cooperative reasoning of the intelligent wireless network nodes, so that the communication time delay is reduced, and the timeliness of model task processing is improved.
In an exemplary embodiment, as shown in fig. 4, determining an inference node in the set of intelligent wireless network nodes that meets the joint inference requirement in step S302 may include:
step S401, determining node time delay requirements and node calculation force requirements according to the joint reasoning requirements.
In the step, the wireless intelligent management and arrangement function can determine the node time delay requirement and the node calculation force requirement according to the joint reasoning requirement. The node delay requirement can be a delay requirement for each inference node, the node delay requirement can include a node transmission delay and a node processing delay, the node transmission delay can include a transmission delay of information related to an inference node joint inference task, and the node processing delay can include a processing delay of an inference node processing model joint inference task. The node calculation force demand can be the calculation force demand of each reasoning node, and the calculation force of the nodes can be measured through parameters such as the size of a video memory. The node time delay requirement and the node calculation force requirement can be contained in the joint reasoning requirement, and can also be determined by a wireless intelligent management and arrangement function according to the requirements of the joint reasoning requirement in terms of time delay, calculation force and the like.
Step S402, determining an inference node in the intelligent wireless network node set according to the node time delay requirement and the node calculation power requirement.
In this step, the wireless intelligent management and arrangement function may determine an inference node in the intelligent wireless network node set according to the node delay requirement and the node calculation power requirement. The inference node is one or more intelligent wireless network nodes which meet the node time delay requirement and the node calculation force requirement. The wireless intelligent management and arrangement function can determine an inference node according to node time delay information and node calculation power information of each intelligent wireless network node in the intelligent wireless network node set. The node time delay information may include information indicating node transmission time delay and node processing time delay of the intelligent wireless network node, and the node calculation power information may include parameters for measuring calculation power, such as video memory size of the intelligent wireless network node.
The scheme of the embodiment can accurately distribute the inference nodes meeting the joint inference requirement in the intelligent wireless network node set based on the node time delay requirement and the node calculation force requirement through the wireless intelligent management and arrangement function, and assist the collaborative execution of the model inference task.
In an exemplary embodiment, as shown in fig. 5, determining an inference node in the intelligent wireless network node set according to the node latency requirement and the node computation power requirement in step S402 may include:
Step S501, a preliminary reasoning node subset meeting node time delay requirements and node calculation force requirements is determined in an intelligent wireless network node set.
In this step, the wireless intelligent management and arrangement function may first determine a subset of the preliminary reasoning nodes satisfying the node delay requirement and the node calculation power requirement in the intelligent wireless network node set. The first-choice reasoning node subset is a subset of the intelligent wireless network node set, each intelligent wireless network node (which can be marked as a first-choice reasoning node) in the subset can meet the node time delay requirement and the node calculation force requirement, the first-choice reasoning node subset meeting the node time delay requirement and the node calculation force requirement is selected in the first step, and then the reasoning nodes are further selected in the subsequent steps.
Step S502, determining the inference node in the primary inference node subset according to the resource utilization rate of each primary inference node in the primary inference node subset.
In this step, the wireless intelligent management and arrangement function may obtain a resource utilization rate of each primary selection inference node in the subset of primary selection inference nodes, where the resource utilization rate may include a computing resource utilization rate and/or a communication resource utilization rate, and may determine an inference node in the subset of primary selection inference nodes according to the resource utilization rate of each primary selection inference node. The first-choice inference nodes in the first-choice inference node subset can be one or more first-choice inference nodes meeting the resource utilization condition, for example, the first-choice inference nodes with the computing resource utilization rate smaller than or equal to the computing resource utilization rate threshold value can be selected as the inference nodes, the first-choice inference nodes with the communication resource utilization rate smaller than or equal to the communication resource utilization rate threshold value can also be selected as the inference nodes, and the like, and if the number of the first-choice inference nodes meeting the resource utilization condition comprises a plurality of first-choice inference nodes, one of the first-choice inference nodes can be randomly selected as the inference node.
The scheme of the embodiment can further consider the resource utilization rate of the nodes on the basis of the node time delay requirement and the node calculation force requirement, and ensure the integral reasoning performance of the intelligent wireless network node set.
In an exemplary embodiment, determining the inference node in the subset of preliminary inference nodes according to the resource utilization of each preliminary inference node in the subset of preliminary inference nodes in step S502 may include:
The method comprises the steps of obtaining an optimized transmission path corresponding to each primary selection inference node in a primary selection inference node subset, obtaining optimized node time delay information of each primary selection inference node according to the optimized transmission path, and determining an inference node in the primary selection inference node subset according to the optimized node time delay information and the resource utilization rate of each primary selection inference node in the primary selection inference node subset.
In this embodiment, after determining a subset of initially selected inference nodes satisfying the node delay requirement and the node calculation requirement in the intelligent wireless network node set, the wireless intelligent management and arrangement function may optimize a transmission path of each initially selected inference node in the intelligent wireless network node set, respectively, to obtain an optimized transmission path corresponding to each initially selected inference node. The transmission path can comprise a path for the initially selected reasoning node to receive and send the relevant information of the model joint reasoning task. The wireless intelligent management and arrangement function can optimize the transmission path of each primary selection inference node to obtain an optimized transmission path, and obtain the optimized node time delay information (namely the optimized node time delay information) of each primary selection inference node according to the optimized transmission path. That is, in this embodiment, the wireless intelligent management and arrangement function further considers the node delay information after optimizing the transmission path of each initially selected inference node, so as to further improve the accuracy of assigning the inference nodes and improve the processing timeliness of the model inference task.
Therefore, in this embodiment, the wireless intelligent management and arrangement function may determine the inference node in the subset of primary inference nodes according to the optimized node delay information and the resource utilization rate of each primary inference node in the subset of primary inference nodes. As an example, the wireless intelligent management and arrangement function may first select, according to the optimized node delay information, a plurality of initially selected inference nodes with the minimum optimized node delay from the initially selected inference node subset, and then select, according to the resource utilization, one or more initially selected inference nodes with the minimum resource utilization from the initially selected inference node subset as the inference nodes.
The scheme of the embodiment can further consider optimizing the transmission path of the node on the basis of the node time delay requirement and the node calculation power requirement, so that the inference node is determined according to the optimized node time delay information and the resource utilization rate, the accuracy of distributing the inference node and the processing timeliness of the model inference task are further improved, and the overall inference performance of the intelligent wireless network node set can also be improved.
In an exemplary embodiment, determining the inference node in the subset of preliminary inference nodes according to the resource utilization of each preliminary inference node in the subset of preliminary inference nodes in step S502 may include:
And determining the inference nodes in the primary inference node subset according to the balance of the resource utilization rate of each primary inference node.
In this embodiment, the wireless intelligent management and arrangement function may determine the inference node in the subset of primary inference nodes according to the balance of the resource utilization rate of each primary inference node in the subset of primary inference nodes, so as to balance the resource utilization rate of each primary inference node and balance the overall inference performance of the intelligent wireless network node set. As an example, the wireless intelligent management and orchestration function may prioritize the first-choice inference node with the lowest resource utilization as the inference node.
In an exemplary embodiment, determining the inference node in the intelligent wireless network node set according to the node latency requirement and the node computation power requirement in step S402 may include:
the method comprises the steps of obtaining preliminary node time delay information and node calculation force information of each intelligent wireless network node in an intelligent wireless network node set, determining an intelligent wireless network node subset meeting node time delay requirements in the intelligent wireless network node set according to the preliminary node time delay information, determining a primary reasoning node subset meeting the node calculation force requirements in the intelligent wireless network node subset according to the node calculation force information, and obtaining reasoning nodes according to the primary reasoning node subset.
The scheme of the embodiment can determine the inference node meeting the node time delay requirement and the node calculation force requirement. The wireless intelligent management and arrangement function can firstly acquire preliminary node time delay information and node calculation force information of each intelligent wireless network node in the intelligent wireless network node set. The preliminary node delay information may be delay information of the intelligent wireless network node before the transmission path of the intelligent wireless network node is optimized. Then, the wireless intelligent management and arrangement function can determine an intelligent wireless network node subset meeting the node delay requirement from the intelligent wireless network node set according to the preliminary node delay information, wherein the intelligent wireless network node subset can contain a plurality of intelligent wireless network nodes meeting the node delay requirement. The wireless intelligent management and arrangement function can then determine a primary selection inference node subset meeting the node calculation force demand from the intelligent wireless network node subset according to the node calculation force information, wherein the primary selection inference node subset can comprise a plurality of intelligent wireless network nodes (which can be recorded as primary selection inference nodes) meeting the node delay demand and the node calculation force demand. Finally, the wireless intelligent management and arrangement function can obtain the inference nodes according to the first-choice inference node subset. In some embodiments, the wireless intelligent management and arrangement function may determine the inference node in the subset of the initially selected inference nodes according to the resource utilization rate of each of the initially selected inference nodes in the subset of the initially selected inference nodes, and the specific processing procedure may refer to the foregoing embodiment of the solution, which is not described herein.
In an exemplary embodiment, determining the node latency requirement and the node computation power requirement according to the joint reasoning requirement in step S401 may include:
and determining the node time delay requirement and the node calculation force requirement according to the task description information, the terminal calculation force information and the total time delay requirement in the joint reasoning requirement.
In this embodiment, the wireless intelligent management and arrangement function may determine the node delay requirement and the node calculation power requirement according to the task description information, the terminal calculation power information, and the total delay requirement in the joint reasoning requirement. The task description information may include, among other things, the type of data that needs to be processed by the model, the model that needs to be used, model reasoning tasks, and so on. The terminal calculation force information may include parameters for measuring the calculation force, such as a memory size of the terminal. The total delay requirement may include a delay requirement representing the whole model inference task of the terminal, and the total delay requirement may include a plurality of parts, as shown in fig. 6, and may include, but is not limited to, a processing delay t1 of the terminal itself, a transmission delay t2 of data related to the model joint inference task sent by the terminal to the inference node, a processing delay t3 of the inference node executing the joint inference task, and a transmission delay t4 of data related to the model joint inference task sent by the inference node to the terminal. As an example, the wireless intelligent management and orchestration function may determine a model local inference task and a local processing delay thereof, which need to be executed by the terminal, according to the task description information, the terminal computing power information, and the total delay requirement, obtain a node delay requirement according to the total delay requirement and the local processing delay, and determine the node computing power requirement according to the node delay requirement, the model local inference task, and the model inference task.
In an exemplary embodiment, after the obtaining of the model joint reasoning request sent by the access node in step S301, the method may further include the following steps:
and the model local reasoning task is determined according to the task description information, the terminal calculation force information and the total time delay requirement in the joint reasoning requirement, and is sent to the terminal through the access node.
In the embodiment, the wireless intelligent management and arrangement function can also determine the local reasoning task of the model of the terminal, so that the accuracy and efficiency of the reasoning task division are improved. The model local reasoning task refers to a part of the model reasoning task, which is executed by the terminal. The wireless intelligent management and arrangement function can determine a model local reasoning task which can be executed by the terminal in the model reasoning task according to task description information, terminal calculation force information and total time delay requirement in the joint reasoning requirement, for example, based on the terminal calculation force information, a part of the model reasoning task, of which the calculation force required by the model reasoning task is smaller than or equal to a calculation force threshold value, can be determined as the model local reasoning task according to the total time delay requirement, and the model local reasoning task can be executed by using the proper calculation force of the terminal so as to meet the total time delay requirement of collaborative reasoning. Therefore, the wireless intelligent management and arrangement function can send the local model reasoning task of the terminal to the terminal through the access node, and the local model reasoning task is used for the terminal to execute so as to obtain a terminal reasoning result.
In an exemplary embodiment, the inference node may be configured to obtain, via the access node, model input data provided by the terminal, and obtain a node inference result according to the model joint inference task and the model input data.
In this embodiment, the inference node may obtain a model joint inference task from the wireless intelligent management and orchestration function, and may obtain model input data provided by the terminal from the access node, where the model input data is model input data required by the inference node to execute the model joint inference task.
The terminal can send the model input data to the access node when sending the model joint reasoning request, can send the model input data to the access node after receiving the model local reasoning task, and can send the model input data to the access node after executing the model local reasoning task. It should be noted that, if the model input data relates to user data (user information), the terminal needs to prompt the user on the interface through the screen of the terminal at least before transmitting the model input data to the access node, and the model input data will be transmitted to the inference node through the access node to process the related information of the processing procedure of the model input data in the method of the present application, and an obvious canceling or authorizing interface component is provided on the interface for the user to cancel the processing or authorizing the processing of the model input data to the model input data, so that the collection, the use and the processing of the related data meet the related regulations.
The model input data may or may not include terminal reasoning results of the terminal. In this regard, the execution of the model joint reasoning task by the reasoning node and the execution of the model local reasoning task by the terminal may have a sequence, or may be parallel execution.
Under the condition of sequence, the terminal can execute the local model reasoning task and then send the terminal reasoning result as the model input data to the access node, at this time, the reasoning node can acquire the model input data (including the terminal reasoning result) provided by the terminal from the access node, execute the model joint reasoning task accordingly, and then transmit the node reasoning result to the terminal through the access node. The inference node may perform the model joint inference task first, and at this time, the terminal sends the model input data (excluding the terminal inference result) to the access node, and the inference node performs the model joint inference task accordingly, and then transmits the node inference result to the terminal via the access node, so that the terminal may perform the model local inference task accordingly.
In the case of parallel execution, the terminal may also send model input data (excluding the terminal's terminal reasoning result) to the access node, from which the reasoning node performs a model joint reasoning task, and then transmit the node reasoning result to the terminal via the access node, during which the terminal may perform the model local reasoning task together.
In a specific implementation, the model corresponding to the model reasoning task for reasoning the node and the terminal required collaborative reasoning can comprise two parts, wherein one part is a terminal model part executed by the terminal, and the second part is a node model part executed by the node. In some embodiments, models corresponding to the model reasoning tasks of the collaborative reasoning needed by the reasoning nodes and the terminals can be deployed in the reasoning nodes and the terminals in advance, and in this way, the reasoning nodes and the terminals can process the corresponding model parts according to the respective reasoning tasks. In some embodiments, the model corresponding to the model reasoning task of the reasoning node and the model of the collaborative reasoning needed by the terminal is deployed on the terminal in advance, and for this purpose, the terminal may transmit the node model part (or the data used for constructing the node model part) to the access node, and the reasoning node may acquire the node model part from the access node (or acquire the data used for constructing the node model part and construct the node model part according to this), so that the reasoning node and the terminal may process the corresponding model part according to the respective reasoning task.
In one exemplary embodiment, the terminal may be configured to obtain a terminal reasoning result according to the node reasoning result and a model local reasoning task of the terminal, and obtain a model reasoning result according to the terminal reasoning result.
In this embodiment, the terminal may obtain the node inference result obtained by the inference node from the access node, and then obtain the terminal inference result according to the node inference result and the model local inference task, and use the terminal inference result as the model inference result of the model inference task. Therefore, the model reasoning results of the model reasoning tasks can be obtained by firstly executing part of model reasoning tasks by the reasoning nodes and then executing part of model reasoning tasks by the terminal according to the node reasoning results.
In an exemplary embodiment, the terminal may be configured to obtain a model inference result based on the node inference result and the terminal inference result.
In this embodiment, the terminal may obtain the node inference result obtained by the inference node from the access node, and the terminal may also perform a model local inference task to obtain a terminal inference result, and then the terminal may obtain the model inference result according to the node inference result and the terminal inference result, for example, the terminal may splice the node inference result and the terminal inference result to obtain the model inference result. Therefore, the model reasoning result can be obtained by the fact that the reasoning nodes and the terminals execute corresponding part of model reasoning tasks in parallel and then the terminals splice the node reasoning result and the terminal reasoning result.
In an exemplary embodiment, the terminal may be configured to send a terminal reasoning result to the access node, so that the reasoning node obtains a node reasoning result according to the model joint reasoning task and the terminal reasoning result, and the terminal is further configured to obtain a model reasoning result according to the node reasoning result.
In this embodiment, the terminal may execute the model local reasoning task to obtain the terminal reasoning result, then the terminal may send the terminal reasoning result to the access node, the reasoning node executes the model joint reasoning task according to the terminal reasoning result to obtain the node reasoning result and sends the node reasoning result to the terminal through the access node, the terminal may obtain the node reasoning result from the access node, and then the node reasoning result may be used as the model reasoning result. Therefore, the terminal can execute part of model reasoning tasks, then the reasoning node executes part of model reasoning tasks according to the terminal reasoning results to obtain node reasoning results, and then the terminal obtains model reasoning results according to the node reasoning results.
In an exemplary embodiment, as shown in fig. 7, there is also provided a model task processing method, which may include the steps of:
in step S701, the terminal sends a model joint reasoning request to the access node.
In step S702, the access node forwards the model joint reasoning request to the general computing resource scheduling module of the wireless intelligent management and arrangement function.
In step S703, the general computing resource scheduling module determines an inference node in the intelligent wireless network node set that meets the joint inference requirement.
And step S704, the general computing resource scheduling module triggers the reasoning node to execute the model joint reasoning task corresponding to the model joint reasoning request.
Step S705, the general calculation resource scheduling module informs the terminal of uploading model input data through the access node.
In step S706, the terminal transmits the model input data to the access node.
In step S707, the inference node obtains model input data from the access node, and executes a model joint inference task.
Step S708, the reasoning node sends the node reasoning result to the terminal through the access node.
Step S709, the terminal obtains a model reasoning result.
In this embodiment, when the terminal faces a complex artificial intelligence model reasoning task, the terminal may encapsulate task description information, terminal calculation information and total delay requirements into a model joint reasoning request because its calculation power is limited and cannot meet all reasoning requirements. Based on the model joint reasoning request, the model reasoning task can be divided into a part executed by the reasoning node and a part executed by the terminal by the general calculation resource scheduling module, and the division basis can comprise task description information, terminal calculation force information, total time delay requirement and the like. The general calculation resource scheduling module can determine a sub-set of primary selection inference nodes in the intelligent wireless network node set according to node time delay requirements and node calculation force requirements, the general calculation resource scheduling module can further optimize transmission paths corresponding to each primary selection inference node in the sub-set of primary selection inference nodes, obtain optimized node time delay information of each primary selection inference node according to the optimized transmission paths, and then the general calculation resource scheduling module can determine inference nodes in the sub-set of primary selection inference nodes according to the optimized node time delay information of each primary selection inference node and the resource utilization rate of each primary selection inference node, for example, the primary selection inference node with low time delay (which can be judged by a time delay threshold) and low resource utilization rate (which can be judged by a resource utilization rate threshold) can be selected as the inference node. The general computing resource scheduling module can send the model joint reasoning task to the reasoning node, and trigger the reasoning node to execute the model joint reasoning task. The general computing resource scheduling module can also inform the terminal to upload the model input data to the access node through the access node so that the inference node executes the model joint inference task. The general computing resource scheduling module may also notify the terminal via the access node to upload the node model part (or data for constructing the node model part) to the access node, and the inference node may obtain the node model part from the access node (or obtain the data for constructing the node model part and construct the node model part accordingly). After the reasoning node finishes the reasoning, the node reasoning result can be sent to the terminal through the access node, and the terminal can finally obtain the model reasoning result according to the node reasoning result and the terminal reasoning result.
According to the scheme, the computing power resources in the intelligent wireless network node can be fully utilized, the reasoning performance is improved, the intelligent wireless network node can provide strong computing power in the intelligent wireless network, the computing power of the intelligent wireless network node is directly embedded in the wireless access network, the forwarding path of data in the network is reduced from the architecture, the extra transmission delay caused by relying on edge computing can be avoided, the efficiency of collaborative reasoning can be remarkably improved, the intelligent wireless network node is closer to a low-time-delay service scene with higher real-time requirements, and meanwhile, the strong computing power of the intelligent wireless network node is cooperated with the local reasoning of a terminal, so that the terminal with limited computing power can also complete a task with high complexity. The intelligent wireless network can realize unified scheduling of calculation power and communication resources in the intelligent wireless network by introducing wireless intelligent management and arrangement functions, dynamically allocate model reasoning tasks to proper intelligent wireless network nodes, give consideration to transmission efficiency and reasoning performance, optimize transmission paths of the nodes, improve timeliness and resource utilization rate of overall task processing, and ensure that the terminal can provide low-delay and high-reliability artificial intelligent service in complex scenes.
It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.
Based on the same inventive concept, the embodiment of the application also provides a model task processing device for realizing the above related model task processing method. The implementation of the solution provided by the device is similar to the implementation described in the above method, so the specific limitation of one or more embodiments of the model task processing device provided below may refer to the limitation of the model task processing method hereinabove, and will not be described herein.
In an exemplary embodiment, as shown in fig. 8, a model task processing device is provided, the device 800 may be applied to a wireless intelligent management and orchestration function, and the device 800 may include:
The request acquisition module 801 is configured to acquire a model joint reasoning request sent by an access node, where the access node is an intelligent wireless network node accessed by a terminal, and the model joint reasoning request is sent to the access node by the terminal;
A node determining module 802, configured to determine an inference node in an intelligent wireless network node set that meets the joint inference requirement;
The task triggering module 803 is configured to trigger the inference node to execute a model joint inference task corresponding to the model joint inference request and transmit a node inference result to the terminal through the access node, so that the terminal obtains a model inference result, where the model inference result is obtained based on a terminal inference result of the terminal and the node inference result.
In an exemplary embodiment, a node determination module 802 is configured to determine a node latency requirement and a node computational power requirement according to the joint inference requirement, and determine the inference node in the set of intelligent wireless network nodes according to the node latency requirement and the node computational power requirement.
In an exemplary embodiment, a node determining module 802 is configured to determine a subset of initially selected inference nodes in the set of intelligent wireless network nodes that satisfy the node latency requirement and the node computational power requirement, and determine the inference nodes in the subset of initially selected inference nodes according to a resource utilization of each initially selected inference node in the subset of initially selected inference nodes.
In an exemplary embodiment, a node determining module 802 is configured to obtain an optimized transmission path corresponding to each primary-selection inference node in the subset of primary-selection inference nodes, obtain optimized node delay information of each primary-selection inference node according to the optimized transmission path, and determine the inference node in the subset of primary-selection inference nodes according to the optimized node delay information of each primary-selection inference node in the subset of primary-selection inference nodes and the resource utilization.
In an exemplary embodiment, the node determining module 802 is configured to determine the inference node in the subset of the initially-selected inference nodes according to the balance of the resource utilization of each of the initially-selected inference nodes.
In an exemplary embodiment, a node determining module 802 is configured to obtain preliminary node delay information and node calculation power information of each intelligent wireless network node in the intelligent wireless network node set, determine a subset of intelligent wireless network nodes meeting the node delay requirement in the intelligent wireless network node set according to the preliminary node delay information, determine a subset of preliminary reasoning nodes meeting the node calculation power requirement in the subset of intelligent wireless network nodes according to the node calculation power information, and obtain the reasoning nodes according to the subset of preliminary reasoning nodes.
In an exemplary embodiment, the node determining module 802 is configured to determine the node latency requirement and the node computing power requirement according to task description information, terminal computing power information and total latency requirement in the joint reasoning requirement.
In an exemplary embodiment, the apparatus 800 may further include a task sending module configured to determine a model local reasoning task of the terminal according to task description information, terminal computing power information and total delay requirement in the joint reasoning requirement, send the model local reasoning task to the terminal through the access node, and perform the model local reasoning task by the terminal to obtain the terminal reasoning result.
In an exemplary embodiment, the inference node is configured to obtain, via the access node, model input data provided by the terminal, and obtain the node inference result according to the model joint inference task and the model input data, where the model input data is sent by the terminal to the access node, and the model input data includes or does not include a terminal inference result of the terminal.
In an exemplary embodiment, the terminal is configured to obtain a terminal inference result according to the node inference result and a model local inference task of the terminal, obtain the model inference result according to the terminal inference result, or obtain the model inference result according to the node inference result and the terminal inference result, or send the terminal inference result to the access node, so that the inference node obtains the node inference result according to the model joint inference task and the terminal inference result, and obtain the model inference result according to the node inference result.
The various modules in the model task processing device described above may be implemented in whole or in part by software, hardware, and combinations thereof. The above modules may be embedded in hardware or may be independent of a processor in the communication device, or may be stored in software in a memory in the communication device, so that the processor may call and execute operations corresponding to the above modules.
In one exemplary embodiment, a communication device is provided that may be used as a wireless intelligent management and orchestration function, the internal structure of which may be as shown in fig. 9. The communication device includes a processor, a memory, an input/output interface, and a communication interface. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface is connected to the system bus through the input/output interface. Wherein the processor of the communication device is configured to provide computing and control capabilities. The memory of the communication device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input/output interface of the communication device is used to exchange information between the processor and the external device. The Communication interface of the Communication device is used for performing wired or wireless Communication with an external terminal, and the wireless Communication can be realized through WIFI, a mobile cellular network, near field Communication (NEAR FIELD Communication) or other technologies. The computer program is executed by a processor to implement a model task processing method.
It will be appreciated by those skilled in the art that the structure shown in fig. 9 is merely a block diagram of a portion of the structure associated with the present inventive arrangements and is not limiting of the communication device to which the present inventive arrangements are applied, and that a particular communication device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.
In an embodiment, there is also provided a communication device including a memory and a processor, the memory storing a computer program, the processor implementing the steps of the method embodiments described above when executing the computer program.
In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored which, when executed by a processor, carries out the steps of the method embodiments described above.
In an embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of the method embodiments described above.
It should be noted that, the user information (including but not limited to user equipment information, user personal information, etc.) and the data (including but not limited to data for analysis, stored data, presented data, etc.) related to the present application are both information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data are required to meet the related regulations.
Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in embodiments provided herein may include at least one of non-volatile memory and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (RESISTIVE RANDOM ACCESS MEMORY, reRAM), magneto-resistive Memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (PHASE CHANGE Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in various forms such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc. The databases referred to in the embodiments provided herein may include at least one of a relational database and a non-relational database. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processor referred to in the embodiments provided in the present application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computation, an artificial intelligence (ARTIFICIAL INTELLIGENCE, AI) processor, or the like, but is not limited thereto.
The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the present application.
The foregoing examples illustrate only a few embodiments of the application and are described in detail herein without thereby limiting the scope of the application. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the application, which are all within the scope of the application. Accordingly, the scope of the application should be assessed as that of the appended claims.