Cloud-edge collaborative computing power acceleration optimization method and platform for AIGC

Abstract

The application discloses an AIGC-oriented cloud-edge collaborative computing power acceleration optimization method and platform, relates to the technical field of computing power optimization, and comprises the following steps: obtaining an AIGC subtask set, a corresponding AIGC subtask characteristic parameter set and an AIGC subtask running requirement set; performing computing power scheduling initialization to generate an initial cloud-edge collaborative computing power scheduling scheme; executing the AIGC subtask, collecting feedback information, obtaining a feedback information set, and judging whether the preset performance characteristics are met; if not, optimizing the initial cloud-edge collaborative computing power scheduling scheme to obtain an optimized cloud-edge collaborative computing power scheduling scheme, and performing cloud-edge collaborative computing power acceleration. The application solves the technical problem that, in the prior art, cloud-edge computing power scheduling cannot be dynamically performed according to AIGC subtask characteristics and running requirements, resulting in low resource allocation efficiency and high response time delay, and achieves the technical effects of improving AIGC task execution efficiency, reducing edge response delay and improving computing power resource utilization.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of computing power optimization, in particular to a cloud-edge collaborative computing power acceleration optimization method and platform for AIGC. BACKGROUND

[0002] In AIGC (artificial intelligence generated content) applications, there are usually multi-modal inputs and heterogeneous computing requirements, and the task execution process needs to be completed collaboratively between the cloud and the edge nodes. Due to the significant differences in input modalities, complexities and response time limits of different sub-tasks, if there is a lack of in-depth analysis and dynamic adaptation of task characteristic parameters and running requirements, the cloud-edge computing power scheduling process is likely to use fixed strategies or static rules, which makes it difficult to flexibly allocate resources according to the actual load of the task, resulting in increased delay due to excessive load on some edge nodes, low utilization rate due to idle resources on some nodes, and affected overall task execution efficiency. SUMMARY

[0003] The application provides a cloud-edge collaborative computing power acceleration optimization method and platform for AIGC, which is used to solve the technical problem that the existing technology cannot dynamically schedule cloud-edge computing power according to the characteristics and running requirements of AIGC sub-tasks, resulting in low resource allocation efficiency and high response time.

[0004] In view of the above problems, the application provides a cloud-edge collaborative computing power acceleration optimization method and platform for AIGC.

[0005] In a first aspect, the application provides a cloud-edge collaborative computing power acceleration optimization method for AIGC, which comprises:

[0006] The multi-modal input data of the target AIGC task is received by the user end, the multi-modal input data is analyzed for task splitting, the AIGC sub-task set, the corresponding AIGC sub-task characteristic parameter set and the AIGC sub-task running requirement set are obtained, the cloud-edge collaborative computing power scheduling initialization is performed by the computing power scheduling agent deployed in the cloud according to the AIGC sub-task characteristic parameter set and the AIGC sub-task running requirement set, the initial cloud-edge collaborative computing power scheduling scheme is generated, the initial cloud-edge collaborative computing power scheduling scheme is transmitted to the cloud and distributed to the edge node set, the corresponding AIGC sub-tasks in the AIGC sub-task set are executed, and the task execution feedback information of each edge node is collected to obtain the feedback information set. If the feedback information set does not meet the preset performance characteristics, the initial cloud-edge collaborative computing power scheduling scheme is optimized by the cloud scheduling agent based on the feedback information set to obtain the optimized cloud-edge collaborative computing power scheduling scheme, and the execution of the AIGC sub-task set is accelerated based on the optimized cloud-edge collaborative computing power scheduling scheme.

[0007] In a second aspect of the present application, an AIGC-oriented cloud-edge collaborative computing power acceleration optimization platform is provided, and the platform comprises:

[0008] a task splitting module configured to receive multi-modal input data of a target AIGC task through a user end, analyze the multi-modal input data for task splitting, and obtain an AIGC sub-task set, a corresponding AIGC sub-task feature parameter set, and an AIGC sub-task running requirement set; a computing power scheduling initialization module configured to perform cloud-edge collaborative computing power scheduling initialization according to the AIGC sub-task feature parameter set and the AIGC sub-task running requirement set through a computing power scheduling agent deployed in the cloud end, and generate an initial cloud-edge collaborative computing power scheduling scheme; a feedback information collection module configured to transmit the initial cloud-edge collaborative computing power scheduling scheme to the cloud end and distribute it to an edge node set, execute corresponding AIGC sub-tasks in the AIGC sub-task set, and collect task execution feedback information of each edge node to obtain a feedback information set; a judgment module configured to judge whether the feedback information set meets a preset performance characteristic, and if not, optimize the initial cloud-edge collaborative computing power scheduling scheme based on the feedback information set by the cloud end scheduling agent to obtain an optimized cloud-edge collaborative computing power scheduling scheme; and a computing power acceleration module configured to perform cloud-edge collaborative computing power acceleration on the execution of the AIGC sub-task set based on the optimized cloud-edge collaborative computing power scheduling scheme.

[0009] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0010] The application receives multi-modal input data of a target AIGC task through a user terminal, analyzes the multi-modal input data to perform task splitting, obtains an AIGC subtask set, a corresponding AIGC subtask characteristic parameter set and an AIGC subtask running requirement set; through a computing power scheduling agent deployed in the cloud, cloud-edge collaborative computing power scheduling initialization is performed according to the AIGC subtask characteristic parameter set and the AIGC subtask running requirement set, and an initial cloud-edge collaborative computing power scheduling scheme is generated; the initial cloud-edge collaborative computing power scheduling scheme is transmitted to the cloud and distributed to an edge node set, corresponding AIGC subtasks in the AIGC subtask set are executed, and task execution feedback information of each edge node is collected to obtain a feedback information set; whether the feedback information set meets a preset performance characteristic is judged, if not, the initial cloud-edge collaborative computing power scheduling scheme is optimized based on the feedback information set by the cloud scheduling agent to obtain an optimized cloud-edge collaborative computing power scheduling scheme; and the execution of the AIGC subtask set is accelerated based on the optimized cloud-edge collaborative computing power scheduling scheme. The application solves the technical problems in the prior art that cloud-edge computing power scheduling cannot be dynamically performed according to AIGC subtask characteristics and running requirements, resulting in low resource allocation efficiency and high response time delay, and through the construction of a collaborative computing power scheduling optimization mechanism based on multi-modal task characteristic analysis and feedback driving, the technical effects of improving AIGC task execution efficiency, reducing edge response delay and improving computing power resource utilization are achieved. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0012] Figure 1 A cloud-edge collaborative computing power acceleration optimization method flowchart for AIGC provided by the embodiment of the present application;

[0013] Figure 2 A cloud-edge collaborative computing power acceleration optimization platform structure diagram for AIGC provided by the embodiment of the present application.

[0014] Legend: task splitting module 11, computing power scheduling initialization module 12, feedback information collection module 13, judgment module 14, computing power acceleration module 15. DETAILED DESCRIPTION

[0015] The application provides an AIGC-oriented cloud-edge collaborative computing power acceleration optimization method and platform, which solves the technical problem that the existing technology cannot dynamically schedule cloud-edge computing power according to AIGC subtask characteristics and running requirements, resulting in low resource allocation efficiency and high response delay. By constructing a collaborative computing power scheduling optimization mechanism based on multi-modal task feature analysis and feedback driving, the technical effects of improving AIGC task execution efficiency, reducing edge response delay, and improving computing power resource utilization are achieved.

[0016] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the application.

[0017] It should be noted that any variation of the terms "comprise" and "have" is intended to cover non-exclusive inclusion, for example, a process, method, platform, product or server comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules that are not clearly listed or inherent to the process, method, product or device.

[0018] Embodiment one, as shown in the application provides an AIGC-oriented cloud-edge collaborative computing power acceleration optimization method, which comprises: Figure 1

[0019] Step S100: receiving multi-modal input data of a target AIGC task through a user terminal, analyzing the multi-modal input data for task splitting, obtaining an AIGC subtask set and corresponding AIGC subtask feature parameter set and AIGC subtask running requirement set.

[0020] In the embodiments of the application, first, the multi-modal input data of the target AIGC task is received through the user terminal. The multi-modal input data refers to information content containing multiple modalities such as text, image, voice, video or three-dimensional model at the same time, which is used to express the comprehensive demand of the user on the generative task.

[0021] Next, the above multi-modal input data is analyzed, a pre-trained multi-modal understanding device is called to perform modal recognition, semantic extraction and function classification, the original AIGC task is split according to different processing intentions and execution logic, and an AIGC subtask set is formed.

[0022] ​Subsequently, the corresponding AIGC subtask feature parameter set of each AIGC subtask in the AIGC subtask set is extracted, which contains key parameters for describing the task calculation complexity and resource consumption demand, such as text input length, image resolution, 3D modeling complexity and response time limit.

[0023] Finally, based on the extracted subtask feature parameters, the AIGC subtask running requirement set of each subtask is parsed and obtained in combination with the built-in rule library.

[0024] Further, the method provided by the application embodiment further comprises:

[0025] The AIGC subtask feature parameters in the AIGC subtask feature parameter set at least include text input length, image resolution, 3D modeling complexity and response time limit.

[0026] In the application embodiment, the AIGC subtask feature parameters in the AIGC subtask feature parameter set at least include text input length, image resolution, 3D modeling complexity and response time limit. The text input length reflects the size of the text data required to be processed by the text task, which affects the calculation complexity of the language model; the image resolution represents the pixel size of the input image in the image task, and high resolution usually accompanies higher memory and computing resource consumption; the 3D modeling complexity is used to measure the fineness of the model structure and the complexity of the physical simulation in the three-dimensional generation task, which directly affects the strength of the graph rendering or space calculation; the response time limit defines the time constraint that the task must complete, which is used to guide the priority allocation and scheduling path selection of the computing power resource.

[0027] Further, in the method provided by the application embodiment, the multi-modal input data of the target AIGC task is received by the user end, the multi-modal input data is parsed for task splitting, the AIGC subtask set, the corresponding AIGC subtask feature parameter set and the AIGC subtask running requirement set are obtained, and the method further comprises:

[0028] The multi-modal input data is subjected to intent recognition by using a pre-trained multi-modal understanding device to extract the AIGC subtask set; the AIGC subtask feature parameter set is obtained by performing subtask feature parameter extraction on the multi-modal input data based on the AIGC subtask set; and the AIGC subtask running requirement set is determined by performing running requirement analysis according to the AIGC subtask feature parameter set.

[0029] In the embodiment of the present application, first, the pre-trained multi-modal understanding device is used for intent recognition on the multi-modal input data. In this process, the multi-modal input data is input into the multi-modal understanding device for semantic analysis and task recognition. The multi-modal understanding device performs intent recognition operation by identifying the target keywords in the input, combining the semantic relationship between the context modalities, and splitting the complete task into several functionally independent sub-tasks, and outputs an AIGC sub-task set.

[0030] Then, for the extracted AIGC sub-task set, the original multi-modal input data is executed for feature parameter extraction operation, and the AIGC sub-task feature parameter set of each sub-task is constructed. For text sub-tasks, the number of characters or Token of the text is counted as the text input length; for image sub-tasks, the width and height of the image are read, and the total number of pixels is calculated as the image resolution; for three-dimensional modeling sub-tasks, the total number of patches recorded in the 3D model file is read, and the number and resolution of texture files are counted to represent the 3D modeling complexity, such as the more patches or the larger texture file, the higher the complexity. For the extraction of response time limit, first check whether the user input contains time requirement parameters, if yes, directly as the response time limit; if not, set a default threshold according to the task type, for example, set 200ms for dialogue task and 5 seconds for image generation task, to ensure that the task execution meets the real-time requirement of the scene.

[0031] Finally, according to the AIGC sub-task feature parameter set, the running requirement analysis is performed, and the parameter-demand table matching method is adopted to find the resource requirement configuration corresponding to each parameter in the preset parameter-demand rule library. For example, the task with image resolution greater than 2K is matched to a high-bandwidth path, and the task with response time limit less than 100ms is matched to a low-delay node. All matching results are integrated to output the corresponding AIGC sub-task running requirement set, including computing resource type, bandwidth demand, delay limit, etc.

[0032] Further, the method provided by the embodiment of the present application further comprises:

[0033] Obtaining a plurality of sample multi-modal input data and a plurality of sample AIGC sub-task sets corresponding to the plurality of sample multi-modal input data, and performing supervised training on the framework based on the feedforward neural network to obtain a pre-trained multi-modal understanding device.

[0034] In the embodiments of the present application, first, a plurality of sample multi-modal input data and a plurality of corresponding sample AIGC sub-task sets are obtained as a data set required for supervised training. The multi-modal input data includes text, images, speech, video and other forms in real or simulated scenes, which are used to comprehensively cover the AIGC task requests that the user may initiate in actual application; the corresponding AIGC sub-task set is artificially annotated, which is used to represent the functional sub-task labels that each type of input data should be split into, such as text generation, image generation, speech synthesis, etc.

[0035] When constructing the training model, a feedforward neural network is used as the basic architecture. The network includes an input layer, a plurality of hidden layers and an output layer, and the input features are mapped to a high dimension through a nonlinear activation function. In the training process, the multi-modal input data is input into the network model after being encoded into a unified mode (such as converting images into feature vectors and encoding text into Token vectors), and compared with the corresponding AIGC sub-task set at the output end. The error is calculated by a cross-entropy loss function, and the network parameters are iteratively updated by a backpropagation algorithm. With the continuous input of samples and the increase of training rounds, the model gradually learns the mapping relationship between multi-modal features and sub-task labels. Finally, after training is completed, a pre-trained multi-modal understandinger is obtained.

[0036] Step S200: Through the computing power scheduling agent deployed in the cloud, cloud-edge collaborative computing power scheduling initialization is performed according to the AIGC sub-task feature parameter set and the AIGC sub-task running requirement set, and an initial cloud-edge collaborative computing power scheduling scheme is generated.

[0037] In the embodiments of the present application, through the computing power scheduling agent deployed in the cloud, first, a resource monitoring interface is called to obtain the real-time resource status of the edge node set, including the GPU model, available video memory, network bandwidth and response time of each node and other information. Then the scheduling agent reads the AIGC sub-task feature parameter set and the AIGC sub-task running requirement set as the scheduling input.

[0038] By using the method of conditional matching, the running requirements of each AIGC sub-task are compared with the resource status of the edge node one by one. For the nodes that meet the required GPU type, memory capacity, bandwidth requirement and response time limit of the task, the nodes are marked as available nodes. The scheduling agent selects the node with the most idle resources or the lightest load as the target execution node among all available nodes.

[0039] The scheduling agent allocates the corresponding target edge node for each sub-task, and records the task number allocated on each node, the GPU type used and the resource allocation. After completing the allocation of all sub-tasks, the initial cloud-edge collaborative computing power scheduling scheme is generated, which includes the correspondence between the tasks and the nodes and the computing power allocation result.

[0040] Step S300: transmitting the initial cloud-edge collaborative computing power scheduling scheme to the cloud end and distributing it to the edge node set, executing the corresponding AIGC subtask in the AIGC subtask set, and collecting the task execution feedback information of each edge node to obtain a feedback information set.

[0041] In the embodiments of the present application, the initial cloud-edge collaborative computing power scheduling scheme is transmitted to the cloud end, and the cloud end distributes each task content specified in the initial cloud-edge collaborative computing power scheduling scheme to the corresponding edge node set, and each edge node executes the target subtask in the AIGC subtask set according to the allocation content. During task execution, the task execution state of each edge node in the preset window is analyzed according to the preset edge node feedback index, the running data including the GPU usage rate sequence, the inference delay sequence, the error rate sequence and the task response time sequence are collected, and each feedback feature is extracted based on the data. Then, the feedback features are combined and interacted to form a task execution feedback feature combination set and a task execution feedback interaction feature set, and the first feedback information of a single node is generated by calculating the mean value of the interaction feature set. Finally, all edge nodes are traversed, and the feedback results of each node are integrated to form a feedback information set.

[0042] Further, in the method provided by the application, the initial cloud-edge collaborative computing power scheduling scheme is transmitted to the cloud end and distributed to the edge node set, the corresponding AIGC subtask in the AIGC subtask set is executed, and the task execution feedback information of each edge node is collected to obtain a feedback information set, which further includes:

[0043] According to the preset edge node feedback index, the task execution state of the first edge node in the edge node set in the preset window is analyzed to obtain the GPU usage rate sequence, the inference delay sequence, the error rate sequence and the task response time sequence; the GPU usage rate sequence, the inference delay sequence, the error rate sequence and the task response time sequence are traversed for multi-head attention analysis to determine the GPU usage rate feedback feature, the inference delay feedback feature, the error rate feedback feature and the task response time feedback feature; the GPU usage rate feedback feature, the inference delay feedback feature, the error rate feedback feature and the task response time feedback feature are combined in pairs to determine a task execution feedback feature combination set; the task execution feedback feature combination set is traversed for feature interaction to determine a task execution feedback interaction feature set; the mean value of the task execution feedback interaction feature set is calculated to determine the first feedback information; the task execution state in the preset window is analyzed according to the preset edge node feedback index to determine the feedback information set by traversing the edge node set.

[0044] In the embodiment of the present application, first, the task execution state of any one edge node (denoted as a first edge node) in the preset window in the edge node set is collected and analyzed according to the preset edge node feedback index. A fixed interval polling sampling method is adopted, and 100 milliseconds is taken as a sampling period to collect GPU utilization, model inference time consumption, abnormal execution flag and response time length from the log of the first edge node to form a GPU utilization rate sequence (recording the percentage of GPU resource occupation in the task execution process), an inference delay sequence (recording the time consumed by each round of model inference), an error rate sequence (recording the proportion of abnormal output or failed tasks in unit time), and a task response time sequence (recording the total time length of each task from receiving to output completion).

[0045] Subsequently, the GPU utilization rate sequence, the inference delay sequence, the error rate sequence and the task response time sequence are traversed, and multi-head attention analysis is performed on each sequence respectively. Specifically, first, each sequence is vectorized, a uniform length T (such as 50 time slices) is set, and the original value of each time point is mapped to a fixed-dimensional vector (such as 64 dimensions) through linear transformation to form an input matrix of T x 64. Subsequently, a plurality of parallel attention heads are constructed on each sequence, each attention head uses different linear weight matrices to generate Query, Key and Value vectors respectively, and the attention weight matrix is calculated through dot product to identify the dependence degree of each time slice in the sequence to other time slices. For example, in the GPU utilization rate sequence [70, 75, 90, 85, 65, 60], attention head A may assign higher weight to position 3 (value 90) to focus on local peak value; attention head B may increase the weight of positions 1 and 2 (values 70 and 75) to identify the rising trend; and attention head C may emphasize positions 5 and 6 (values 65 and 60) to capture the rapid decline feature. The output of each attention head is a weighted feature sequence with the same length as the input. The outputs of multiple attention heads are spliced to form a matrix of T x (N x d) (N is the number of attention heads, and d is the output dimension of each head), and then reduced to a single vector representation through subsequent linear transformation. The vector is a feature representation fused from multiple local perspectives, which is the final feedback feature output. After this process, the GPU utilization rate sequence generates a GPU utilization rate feedback feature, the inference delay sequence generates an inference delay feedback feature, the error rate sequence generates an error rate feedback feature, and the task response time sequence generates a task response time feedback feature.

[0046] Then, Cartesian product method is performed on the above four feedback features to generate a total of 6 groups of feature pairs in pairs, and a task execution feedback feature combination set is constructed, each group containing two feedback features of different dimensions, such as GPU utilization rate and inference delay, or error rate and response time.

[0047] Subsequently, the task execution feedback feature combination set is traversed, and interaction analysis is performed on each combination. Specifically, feature similarity analysis is performed on each combination, the relative change trend of the features in multiple time segments is extracted to form a feature similarity set, and then the proportion of the similarity values in the similarity set that are greater than or equal to a preset feature similarity threshold is counted. If the proportion meets a preset proportion condition, the combination is interacted with to generate a corresponding task execution feedback interaction feature; if not, the current combination is skipped, and the next combination is analyzed until all combinations are traversed and a task execution feedback interaction feature set is output.

[0048] After the task execution feedback interaction feature set is constructed, an arithmetic average operation is performed on all feature values in the task execution feedback interaction feature set to obtain a comprehensive performance result of the first edge node in the current time window, which is determined as the first feedback information.

[0049] Finally, the edge node set is traversed, and the same collection, feature extraction, feature combination, and interaction analysis operations as described above are performed on each edge node to generate feedback information of each node, which is integrated to form a feedback information set.

[0050] Further, in the method provided by the application embodiment, the task execution feedback feature combination set is traversed to perform feature interaction to determine a task execution feedback interaction feature set, and the method further includes:

[0051] The first task execution feedback feature combination in the task execution feedback feature combination set is extracted to perform combination-internal feature similarity analysis to determine a first combination-internal feature similarity set; whether the proportion of the feature similarities in the first combination-internal feature similarity set that are greater than or equal to a preset feature similarity meets a preset proportion is counted, and if so, the first task execution feedback feature combination is interacted with based on the first combination-internal feature similarity set to obtain a first task execution feedback interaction feature; if not, a second task execution feedback feature combination is extracted again from the task execution feedback feature combination set for analysis until the task execution feedback feature combination set is analyzed completely to obtain the task execution feedback interaction feature set.

[0052] In the application embodiment, a random first task execution feedback feature combination, such as a GPU usage rate feedback feature and an inference delay feedback feature, is first extracted from the task execution feedback feature combination set. Combination-internal feature similarity analysis is performed on the first task execution feedback feature combination, a cosine similarity analysis method is used to evaluate the change consistency between the two feedback features point by point at each time segment position to generate a set of time sequence similarity values, which form a first combination-internal feature similarity set.

[0053] Then, the first combination of the feature similarity set is statistically processed, and the interval matching counting method is used to count the number of similarity greater than or equal to the preset feature similarity threshold (such as 0.8) in the set, and calculate the proportion of the total length of the set. If the proportion is greater than or equal to the preset proportion (such as 80%), it indicates that the current two feedback features show high consistency in most time slices, and it can be considered that they have a stable correlation.

[0054] At this time, the feature interaction is performed on the first task execution feedback feature combination based on the first combination of the feature similarity set. In this process, first, the first combination of the feature similarity set is normalized to scale the similarity value to a unified interval, and a combination adjacency matrix is constructed. Then, the combination adjacency matrix is used to interact and enhance the two task execution feedback features. Finally, the mean value processing operation is performed on the result of the interaction and enhancement to obtain the result representing the overall interaction behavior of the feature combination, that is, the first task execution feedback interaction feature.

[0055] If the above proportion does not meet the preset requirement, the current feature combination is skipped, and the second task execution feedback feature combination is extracted from the task execution feedback feature combination set, and the above feature similarity analysis, proportion judgment and feature interaction process are repeated. All feature combinations in the task execution feedback feature combination set are traversed in turn, only the combinations that meet the conditions are interacted, and the corresponding task execution feedback interaction features are generated one by one.

[0056] Finally, the interaction feature results of all valid combinations are summarized to construct and output the task execution feedback interaction feature set.

[0057] Further, in the method provided by the application embodiment, the feature interaction is performed on the first task execution feedback feature combination based on the first combination of the feature similarity set to obtain the first task execution feedback interaction feature, and the method further comprises:

[0058] The first combination of the feature similarity set is normalized to construct a combination adjacency matrix, and the task execution feedback features in the first task execution feedback feature combination are interacted and enhanced by using the combination adjacency matrix, and the interaction and enhancement result is processed by mean value to determine the first task execution feedback interaction feature.

[0059] In the application embodiment, first, the normalization processing is performed on the first combination of the feature similarity set, and the minimum-maximum scaling method is used to compress the similarity value between each pair of feedback features to the [0, 1] interval to obtain the combination of the feature similarity set. Then, the combination adjacency matrix is constructed based on the combination of the feature similarity set. The matrix is a symmetric matrix, and the rows and columns represent the respective task execution feedback features in the first task execution feedback feature combination, and each matrix element represents the normalized similarity between the corresponding two features.

[0060] Then, the task execution feedback features are taken as nodes, and the combined adjacency matrix is taken as edge weight graph structure to construct a graph structure model and perform graph convolution operation to interactively enhance the feature map. In this process, a numerical sequence representing the original feature vector of each node (i.e., feedback feature) in the graph is initialized, for example, the GPU usage feedback feature is [80, 84, 82]. The graph convolution takes the feature map as input, and uses the connection relationship in the combined adjacency matrix to perform an aggregation calculation on the feature values of the neighbor nodes in the neighborhood of each node. Taking the GPU usage feedback feature node as an example, if the similarity between it and the inference delay feedback feature is 0.9, the value will participate in the feature update operation of the GPU node, thereby forming an enhanced GPU usage feedback feature representation. After performing graph convolution, the interactive enhancement result of each feedback feature node is obtained, that is, a set of enhanced vectors that fuse the context information of the adjacent feedback features. For example, the original GPU usage feedback feature is [80, 84, 82], the inference delay feedback feature is [110, 115, 120], and after the graph convolution operation, the GPU feedback node may be updated to [86.2, 88.5, 87.4], reflecting its fusion understanding of the delay fluctuation trend in task execution. The enhanced results of all nodes together constitute the interactive enhancement result.

[0061] Then, the time sequence vector of each task execution feedback feature in the interactive enhancement result is processed by mean value, and a sliding average method is used to process with a sliding window of three frames. For example, the mean value processing result of the above enhanced GPU usage feedback feature [86.2, 88.5, 87.4] is (86.2+88.5+87.4) / 3=87.37. Similarly, for other features such as the error rate feedback feature [0.03, 0.05, 0.06], the mean value processing result is 0.0467.

[0062] Finally, the four processed feedback feature mean value vectors are spliced in order into an overall feedback feature vector as the first task execution feedback interaction feature output.

[0063] Step S400: Determine whether the feedback information set meets the preset performance characteristics. If not, the cloud scheduling agent optimizes the initial cloud-edge collaborative computing power scheduling scheme based on the feedback information set to obtain an optimized cloud-edge collaborative computing power scheduling scheme.

[0064] In the embodiments of the present application, it is first determined whether the feedback information set meets the preset performance characteristics, and the preset performance characteristics include but are not limited to the following task execution indicators: the average GPU usage is less than 85%, the inference delay is not more than 500ms, the error rate is less than 2%, and the task response time is within the tolerance threshold.

[0065] When the feedback information set does not satisfy the preset performance characteristics, the cloud scheduling agent optimizes the initial cloud-edge collaborative computing power scheduling scheme based on the feedback information set. In this process, the cloud scheduling agent first calculates the deviation degree between the feedback information set and the preset performance characteristics, and constructs a positive deviation degree set and a negative deviation degree set respectively. Then, according to the deviation data, a corresponding positive adjustment coefficient set and a negative adjustment coefficient set are generated. According to the above adjustment coefficients, the computing power allocation of each edge node in the initial cloud-edge collaborative computing power scheduling scheme is adjusted to form a new adjusted cloud-edge collaborative computing power scheduling scheme. This scheme will be adjusted and adapted according to the feedback information set, and the adjustment fitness is obtained. If the adjustment fitness result satisfies the preset fitness, the scheme is confirmed as the optimized cloud-edge collaborative computing power scheduling scheme and applied to the actual task distribution; if it does not satisfy the preset performance characteristics, the adjustment and analysis process will continue to be iteratively executed under the guidance of the current adjustment coefficient until the optimized scheme that satisfies the preset performance characteristics is obtained.

[0066] Further, the method provided by the application embodiment further comprises:

[0067] Step a: calculate the deviation degree of the feedback information set and the preset performance characteristics, and obtain a positive deviation degree set and a negative deviation degree set; step b: analyze based on the positive deviation degree set and the negative deviation degree set, and obtain a positive adjustment coefficient set and a negative adjustment coefficient set; step c: according to the positive adjustment coefficient set and the negative adjustment coefficient set, randomly increase or decrease the computing power allocated to the edge nodes in the edge node set in the initial cloud-edge collaborative computing power scheduling scheme, and obtain an adjusted cloud-edge collaborative computing power scheduling scheme; step d: analyze the adjustment fitness of the adjusted cloud-edge collaborative computing power scheduling scheme in combination with the feedback information set, and determine whether the analysis result meets the requirements; if yes, the adjusted cloud-edge collaborative computing power scheduling scheme is used as the optimized cloud-edge collaborative computing power scheduling scheme; step e: if not, repeat steps c-d until the analysis result meets the requirements.

[0068] In the application embodiment, first, the difference between the feedback information set and the preset performance characteristics is compared, and the difference between each performance index (such as GPU utilization, inference delay, error rate, and task response time) in the feedback data and the corresponding preset performance characteristics is calculated. According to the positive and negative directions of the difference, the positive deviation and the negative deviation are divided respectively, and a positive deviation set and a negative deviation set are formed.

[0069] Then, based on the positive deviation degree set and the negative deviation degree set, analysis is performed, in which any one of the positive deviation degrees in the positive deviation degree set is divided by the sum of the positive deviation degrees in the positive deviation degree set, and 1 is subtracted from the calculation result respectively to obtain a positive adjustment coefficient set. The absolute value of any one of the negative deviation degrees in the negative deviation degree set is divided by the sum of the absolute values of the negative deviation degrees in the negative deviation degree set, and 1 is subtracted from the calculation result respectively to obtain a negative adjustment coefficient set.

[0070] Then, according to the positive adjustment coefficient set and the negative adjustment coefficient set, the allocated computing power of the edge nodes in the initial cloud-edge collaborative computing power scheduling scheme is randomly increased or decreased. In this process, for the edge nodes with positive deviation, the allocated computing power resources of the nodes are appropriately reduced according to the corresponding positive adjustment coefficients to avoid resource redundancy or overload calculation; and for the edge nodes with negative deviation, the allocated computing power resources are appropriately increased in proportion to the negative adjustment coefficients to improve their ability to process tasks. The above computing power adjustment adopts an adjustment method with a random disturbance factor, that is, a normal distribution disturbance within a certain range is introduced on the basis of the adjustment coefficient, which is used to avoid falling into a local optimum, and finally an adjusted cloud-edge collaborative computing power scheduling scheme is generated.

[0071] Next, the adjusted cloud-edge collaborative computing power scheduling scheme is adjusted fitness analysis combined with the feedback information set. In this process, first, the feedback information set and the computing power allocation of each edge node in the adjusted cloud-edge collaborative computing power scheduling scheme are combined to construct input features. The feedback information set includes actual running indicators of each edge node, such as GPU usage, inference delay, task response time, node temperature and power consumption, etc. The adjusted cloud-edge collaborative computing power scheduling scheme contains the allocated computing power parameters of the edge nodes, such as CPU core number, GPU quota and bandwidth resource, etc. The above feedback information set and the adjustment scheme are jointly input into the pre-trained neural network model to obtain the adjustment fitness of the current adjustment scheme. The neural network model is trained based on the historical task running records to construct a training set and complete the training. In the training process, the model input is the combination of the feedback information in the historical tasks and the scheduling parameters at that time, and the output is the adjustment fitness score label corresponding to the combination, which is pre-labeled by technical experts to measure the matching degree of the scheduling scheme to the current node state. The training adopts a supervised learning method, uses mean square error as the loss function for parameter update, and the model can be used for online prediction of adjustment fitness after convergence.

[0072] Then, the adjustment fitness is output as an analysis result, and is compared with a preset fitness. If the adjustment fitness is higher than a preset fitness threshold (for example, 0.85), it is indicated that the current scheme can better match the node running state reflected by the current feedback information set, and the adjustment cloud edge collaborative computing power scheduling scheme is confirmed as the optimized cloud edge collaborative computing power scheduling scheme. If the adjustment fitness output by the model does not reach the threshold, it is determined that the current scheme does not meet the requirement, and the foregoing adjustment process is continued, that is, the allocation of computing power of each node is adjusted according to the positive adjustment coefficient set and the negative adjustment coefficient set, a new adjustment scheme is generated, and adjustment fitness analysis is performed again. The iteration is continued until the adjustment fitness output by the model in a certain round of analysis meets the requirement, and finally the optimized cloud edge collaborative computing power scheduling scheme is obtained.

[0073] Further, the method provided by the application embodiment further comprises:

[0074] The absolute value of each negative deviation degree in the negative deviation degree set is divided by the sum of the absolute values of all deviation degrees in the negative deviation degree set to obtain a corresponding proportion value, and 1 is subtracted from the proportion value to obtain a corresponding negative adjustment coefficient.

[0075] In the application embodiment, to obtain the positive adjustment coefficient set and the negative adjustment coefficient set, for any positive deviation degree value in the positive deviation degree set, the positive deviation degree value is first divided by the sum of the values in the positive deviation degree set to obtain a proportion value in the overall deviation, and 1 is subtracted from the proportion value to obtain a corresponding positive adjustment coefficient. The positive adjustment coefficient set is obtained in this way.

[0076] Similarly, for the negative deviation degree set, the absolute value of each negative deviation degree is divided by the sum of the absolute values of all deviation degrees in the negative deviation degree set to obtain a corresponding proportion value, and 1 is subtracted from the proportion value to obtain a corresponding negative adjustment coefficient. The negative adjustment coefficient set is obtained in this way.

[0077] Step S500: cloud edge collaborative computing power acceleration is performed on the execution of the AIGC subtask set based on the optimized cloud edge collaborative computing power scheduling scheme.

[0078] In the embodiment of the present application, based on the optimized cloud-edge collaborative computing power scheduling scheme, when the cloud-edge collaborative computing power acceleration is performed on the AIGC subtask set execution process, the edge node allocation result in the optimized cloud-edge collaborative computing power scheduling scheme is matched with the subtask characteristic parameters, the computing power distribution is performed according to the subtask running demand and the computing power node state, through the parallel calculation and result feedback mechanism of the edge node and the cloud node, the efficient processing of the AIGC subtask is realized, and the overall inference process is accelerated.

[0079] In the embodiment of the present application, as described above, the embodiment of the present application has at least the following technical effects:

[0080] The present application receives multi-modal input data of a target AIGC task through a user end, analyzes the multi-modal input data for task splitting, obtains an AIGC subtask set, a corresponding AIGC subtask characteristic parameter set and an AIGC subtask running demand set; through a computing power scheduling agent deployed in the cloud, the cloud-edge collaborative computing power scheduling is initialized according to the AIGC subtask characteristic parameter set and the AIGC subtask running demand set, and an initial cloud-edge collaborative computing power scheduling scheme is generated; the initial cloud-edge collaborative computing power scheduling scheme is transmitted to the cloud and distributed to an edge node set, the corresponding AIGC subtask in the AIGC subtask set is executed, and the task execution feedback information of each edge node is collected to obtain a feedback information set; it is judged whether the feedback information set meets the preset performance characteristics, if not, the initial cloud-edge collaborative computing power scheduling scheme is optimized based on the feedback information set by the cloud scheduling agent to obtain an optimized cloud-edge collaborative computing power scheduling scheme; the execution of the AIGC subtask set is accelerated based on the optimized cloud-edge collaborative computing power scheduling scheme. The present application solves the technical problems that the cloud-edge computing power cannot be dynamically scheduled according to the AIGC subtask characteristics and running demand in the prior art, resulting in low resource allocation efficiency and high response time delay. Through the construction of a collaborative computing power scheduling optimization mechanism based on multi-modal task characteristic analysis and feedback driving, the technical effects of improving AIGC task execution efficiency, reducing edge response delay and improving computing power resource utilization are achieved.

[0081] Embodiment two, based on the same inventive concept as the AIGC-oriented cloud-edge collaborative computing power acceleration optimization method in the foregoing embodiments, as shown in Figure 2 The present application provides an AIGC-oriented cloud-edge collaborative computing power acceleration optimization platform, and the platform and method embodiments in the embodiment of the present application are based on the same inventive concept. The platform comprises:

[0082] The task splitting module 11 is configured to receive multi-modal input data of a target AIGC task through a user terminal, parse the multi-modal input data for task splitting, and obtain an AIGC subtask set, a corresponding AIGC subtask feature parameter set, and an AIGC subtask running requirement set; the computing power scheduling initialization module 12 is configured to perform cloud-edge collaborative computing power scheduling initialization according to the AIGC subtask feature parameter set and the AIGC subtask running requirement set through a computing power scheduling agent deployed on a cloud side, and generate an initial cloud-edge collaborative computing power scheduling scheme; the feedback information collection module 13 is configured to transmit the initial cloud-edge collaborative computing power scheduling scheme to the cloud side and distribute it to an edge node set, execute corresponding AIGC subtasks in the AIGC subtask set, and collect task execution feedback information of each edge node to obtain a feedback information set; the judgment module 14 is configured to judge whether the feedback information set meets a preset performance characteristic; if not, the cloud-side scheduling agent optimizes the initial cloud-edge collaborative computing power scheduling scheme based on the feedback information set to obtain an optimized cloud-edge collaborative computing power scheduling scheme; and the computing power acceleration module 15 is configured to perform cloud-edge collaborative computing power acceleration on the execution of the AIGC subtask set based on the optimized cloud-edge collaborative computing power scheduling scheme.

[0083] Further, the platform is further configured to implement the following functions:

[0084] The AIGC subtask feature parameter set includes at least a text input length, an image resolution, a 3D modeling complexity, and a response time limit.

[0085] Further, the platform is further configured to implement the following functions:

[0086] The multi-modal input data is subjected to intent recognition by a pre-trained multi-modal understanding device to extract an AIGC subtask set; the multi-modal input data is subjected to subtask feature parameter extraction based on the AIGC subtask set to obtain an AIGC subtask feature parameter set; and the AIGC subtask feature parameter set is subjected to running requirement analysis to determine an AIGC subtask running requirement set.

[0087] Further, the platform is further configured to implement the following functions:

[0088] A plurality of sample multi-modal input data and a plurality of sample AIGC subtask sets are obtained to supervise training of a framework based on a feedforward neural network to obtain a pre-trained multi-modal understanding device.

[0089] Further, the platform is further configured to implement the following functions:

[0090] According to the preset edge node feedback index, the task execution state of a first edge node in the preset window in the edge node set is analyzed, and a GPU usage rate sequence, an inference delay sequence, an error rate sequence, and a task response time sequence are obtained; multi-head attention analysis is performed on the GPU usage rate sequence, the inference delay sequence, the error rate sequence, and the task response time sequence, and GPU usage rate feedback features, inference delay feedback features, error rate feedback features, and task response time feedback features are determined; the GPU usage rate feedback features, the inference delay feedback features, the error rate feedback features, and the task response time feedback features are combined in pairs to determine a task execution feedback feature combination set; feature interaction is performed on the task execution feedback feature combination set to determine a task execution feedback interaction feature set; the mean value of the task execution feedback interaction feature set is calculated to determine first feedback information; the task execution state in the preset window is analyzed according to the preset edge node feedback index to determine the feedback information set.

[0091] Further, the platform is also used to implement the following functions:

[0092] The first task execution feedback feature combination in the task execution feedback feature combination set is extracted for combination internal feature similarity analysis to determine a first combination internal feature similarity set; whether the proportion of feature similarities greater than or equal to a preset feature similarity in the first combination internal feature similarity set meets a preset proportion is counted, if yes, the first task execution feedback feature combination is combined with the first combination internal feature similarity set for feature interaction to obtain a first task execution feedback interaction feature; if not, the second task execution feedback feature combination is extracted again from the task execution feedback feature combination set for analysis until the task execution feedback feature combination set is analyzed completely to obtain the task execution feedback interaction feature set.

[0093] Further, the platform is also used to implement the following functions:

[0094] The first combination internal feature similarity set is normalized to construct a combination adjacency matrix; the task execution feedback features in the first task execution feedback feature combination are interactively enhanced using the combination adjacency matrix, and the interactive enhancement results are processed by mean value to determine the first task execution feedback interaction feature.

[0095] Further, the platform is also used to implement the following functions:

[0096] Step a: calculate the deviation degree of the feedback information set and the preset performance characteristics, obtain a positive deviation degree set and a negative deviation degree set; step b: based on the positive deviation degree set and the negative deviation degree set, analysis is carried out to obtain a positive adjustment coefficient set and a negative adjustment coefficient set; step c: according to the positive adjustment coefficient set and the negative adjustment coefficient set, the allocated computing power of the edge node in the initial cloud-edge collaborative computing power scheduling scheme is randomly increased or decreased to obtain an adjusted cloud-edge collaborative computing power scheduling scheme; step d: combining the feedback information set, the adjusted cloud-edge collaborative computing power scheduling scheme is adjusted and fitness analyzed, and it is judged whether the analysis result meets the requirements, if yes, the adjusted cloud-edge collaborative computing power scheduling scheme is used as the optimized cloud-edge collaborative computing power scheduling scheme; step e: if not, repeat steps c-d until the analysis result meets the requirements.

[0097] Further, the platform is also used to realize the following functions:

[0098] respectively, and the calculation results are respectively subtracted by 1 to obtain a positive adjustment coefficient set; respectively, the absolute value of any one negative deviation degree in the negative deviation degree set is divided by the absolute value sum of the negative deviation degree set, and the calculation results are respectively subtracted by 1 to obtain a negative adjustment coefficient set.

[0099] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above describes a specific embodiment of the present application. The processes depicted in the drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or may be advantageous.

[0100] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0101] The present application and the drawings are only exemplary description of the present application, and are considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalent technology, the present application intends to include these modifications and changes.

Claims

1. A cloud-edge collaborative computing power acceleration optimization method for AIGC, characterized in that, The method includes: The system receives multimodal input data of the target AIGC task through the user terminal, parses the multimodal input data to split the task, and obtains a set of AIGC subtasks, a set of corresponding AIGC subtask feature parameters, and a set of AIGC subtask running requirements. By deploying a computing power scheduling agent in the cloud, cloud-edge collaborative computing power scheduling is initialized based on the AIGC subtask feature parameter set and AIGC subtask operation requirement set, generating an initial cloud-edge collaborative computing power scheduling scheme. The initial cloud-edge collaborative computing power scheduling scheme is transmitted to the cloud and distributed to the set of edge nodes. The corresponding AIGC subtasks in the AIGC subtask set are executed, and the task execution feedback information of each edge node is collected to obtain the feedback information set. Determine whether the feedback information set meets the preset performance characteristics. If not, the cloud scheduling agent optimizes the initial cloud-edge collaborative computing power scheduling scheme based on the feedback information set to obtain an optimized cloud-edge collaborative computing power scheduling scheme. The execution of the AIGC subtask set is accelerated by cloud-edge collaborative computing power based on the optimized cloud-edge collaborative computing power scheduling scheme. The initial cloud-edge collaborative computing power scheduling scheme is transmitted to the cloud and distributed to the set of edge nodes. The corresponding AIGC subtasks in the AIGC subtask set are executed, and task execution feedback information from each edge node is collected to obtain a set of feedback information, including: The task execution status of the first edge node in the edge node set within a preset window is analyzed according to the preset edge node feedback indicators to obtain the GPU utilization rate sequence, inference latency sequence, error rate sequence and task response time sequence. Multi-head attention analysis is performed by traversing the GPU utilization sequence, inference latency sequence, error rate sequence, and task response time sequence to determine the GPU utilization feedback features, inference latency feedback features, error rate feedback features, and task response time feedback features. The GPU utilization feedback features, inference latency feedback features, error rate feedback features and task response time feedback features are combined in pairs to determine the task execution feedback feature combination set. Traverse the set of task execution feedback feature combinations to perform feature interactions and determine the set of task execution feedback interaction features; Calculate the mean of the task execution feedback interaction feature set to determine the first feedback information; Traverse the set of edge nodes and analyze the task execution status within a preset window based on the preset edge node feedback indicators to determine the set of feedback information. Determine whether the feedback information set meets the preset performance characteristics. If not, the cloud scheduling agent optimizes the initial cloud-edge collaborative computing power scheduling scheme based on the feedback information set to obtain an optimized cloud-edge collaborative computing power scheduling scheme, including: Step a: Calculate the deviation between the feedback information set and the preset performance characteristics to obtain a positive deviation set and a negative deviation set; Step b: Based on the positive deviation set and the negative deviation set, perform analysis to obtain the positive adjustment coefficient set and the negative adjustment coefficient set; Step c: According to the positive adjustment coefficient set and the negative adjustment coefficient set, randomly increase or decrease the computing power allocated to the edge nodes in the edge node set in the initial cloud-edge collaborative computing power scheduling scheme to obtain the adjusted cloud-edge collaborative computing power scheduling scheme. Step d: Combine the feedback information set to perform an adjustment fitness analysis on the adjusted cloud-edge collaborative computing power scheduling scheme, and determine whether the analysis results meet the requirements. If so, the adjusted cloud-edge collaborative computing power scheduling scheme is taken as the optimized cloud-edge collaborative computing power scheduling scheme. Step e: If not, repeat step cd until the analysis results meet the requirements.

2. The cloud-edge collaborative computing power acceleration optimization method for AIGC as described in claim 1, characterized in that, The AIGC subtask feature parameter set includes at least the text input length, image resolution, 3D modeling complexity, and response time.

3. The cloud-edge collaborative computing power acceleration optimization method for AIGC as described in claim 2, characterized in that, The system receives multimodal input data for the target AIGC task from the user terminal, parses the multimodal input data to perform task decomposition, and obtains a set of AIGC subtasks, a corresponding set of AIGC subtask feature parameters, and a set of AIGC subtask execution requirements, including: The multimodal input data is used to perform intent recognition using a pre-trained multimodal understander to extract an AIGC subtask set; Based on the AIGC subtask set, subtask feature parameters are extracted from the multimodal input data to obtain the AIGC subtask feature parameter set. The execution requirements set of the AIGC subtask is determined by parsing the set of characteristic parameters of the AIGC subtask.

4. The cloud-edge collaborative computing power acceleration optimization method for AIGC as described in claim 3, characterized in that, Multiple sample multimodal input data and corresponding multiple sample AIGC subtask sets are obtained to supervise the training of the framework built on the feedforward neural network, and a pre-trained multimodal understander is obtained.

5. The cloud-edge collaborative computing power acceleration optimization method for AIGC as described in claim 1, characterized in that, Traverse the set of task execution feedback feature combinations to perform feature interactions, and determine the set of task execution feedback interaction features, including: Extract the first task execution feedback feature combination from the task execution feedback feature combination set, perform feature similarity analysis within the combination, and determine the first feature similarity set within the combination; If the proportion of feature similarity greater than or equal to a preset feature similarity in the feature similarity set within the first combination meets the preset proportion, then, in conjunction with the feature similarity set within the first combination, feature interaction is performed on the first task execution feedback feature combination to obtain the first task execution feedback interaction feature. If not, continue to extract the second task execution feedback feature combination from the task execution feedback feature combination set for analysis, until the task execution feedback feature combination set is fully analyzed to obtain the task execution feedback interaction feature set.

6. The cloud-edge collaborative computing power acceleration optimization method for AIGC as described in claim 5, characterized in that, Then, combining the feature similarity set within the first combination, feature interaction is performed on the first task execution feedback feature combination to obtain the first task execution feedback interaction features, including: The feature similarity set within the first combination is normalized to construct a combined adjacency matrix; The combined adjacency matrix is ​​used to enhance the interaction of the task execution feedback features within the first task execution feedback feature combination, and the mean of the enhancement results is applied to determine the first task execution feedback interaction features.

7. The cloud-edge collaborative computing power acceleration optimization method for AIGC as described in claim 1, characterized in that, include: Each positive deviation in the positive deviation set is compared to the sum of the positive deviation sets mentioned above, and then the result is subtracted from 1 to obtain the positive adjustment coefficient set. The absolute value of any negative deviation in the negative deviation set is divided by the sum of the absolute values ​​of the negative deviation set, and then subtracted from the result by 1 to obtain the set of negative adjustment coefficients.

8. A cloud-edge collaborative computing power acceleration and optimization platform for AIGC, characterized in that, The platform is used to execute the cloud-edge collaborative computing power acceleration optimization method for AIGC as described in any one of claims 1-7, and the platform includes: The task splitting module is used to receive multimodal input data of the target AIGC task through the user terminal, parse the multimodal input data to split the task, and obtain the AIGC subtask set, the corresponding AIGC subtask feature parameter set, and the AIGC subtask running requirement set. The computing power scheduling initialization module is used to initialize cloud-edge collaborative computing power scheduling based on the AIGC subtask feature parameter set and AIGC subtask operation requirement set through a computing power scheduling agent deployed in the cloud, and generate an initial cloud-edge collaborative computing power scheduling scheme. The feedback information collection module is used to transmit the initial cloud-edge collaborative computing power scheduling scheme to the cloud and distribute it to the set of edge nodes, execute the corresponding AIGC sub-tasks in the AIGC sub-task set, and collect the task execution feedback information of each edge node to obtain the feedback information set. The judgment module is used to determine whether the feedback information set meets the preset performance characteristics. If not, the cloud scheduling agent optimizes the initial cloud-edge collaborative computing power scheduling scheme based on the feedback information set to obtain an optimized cloud-edge collaborative computing power scheduling scheme. The computing power acceleration module is used to accelerate the execution of the AIGC subtask set based on the optimized cloud-edge collaborative computing power scheduling scheme.

Images