A media content generation method, apparatus, and program product

Abstract

This paper discloses a media content generation method, apparatus, and program product, relating to the field of computer applications. It solves the problems of high communication costs and low computational efficiency in media content generation. The method includes: inputting acquired first request content into a first model, which is distributed across multiple computing nodes; in response to a first execution operation, acquiring deployment change information of the first model, which is determined based on first record information generated by the computing nodes; in response to a second execution operation, updating the first model based on the deployment change information and determining a first feature representation sequence through the computing nodes; and generating media content based on the first feature representation sequence. This method significantly reduces the communication costs of media content generation through a distributed network structure, effectively improves the computational efficiency of the first model, and better guarantees the generation effect of the media content.

Description

Technical Field

[0001] This article relates to the field of computer application technology, and in particular to a media content generation method, apparatus, and program product. Background Technology

[0002] Currently, media content generation is mainly achieved using network models, and most of these network models are a combination of diffusion models and expert models. Firstly, the model is large in scale and often needs to be distributed across multiple computing nodes, which then undertake the distributed operation of the network model.

[0003] However, for network models whose network structure is distributed across multiple computing nodes and which include diffusion models and expert models, the implementation of media content generation will generate more communication overhead due to the operating characteristics of diffusion models and expert models. It will also suffer from the problem of uneven load on computing nodes as in the first model. This will lead to increased communication costs and reduced computing efficiency, thereby affecting the effect of media content generation. Summary of the Invention

[0004] This paper provides a method, apparatus, and program product for generating media content, which solves the problems of high communication costs and low computational efficiency in media content generation.

[0005] Firstly, this paper provides a method for generating media content, which includes: Obtain the first request content and input the first request content into the first model, which is distributed across multiple computing nodes; In response to the first execution operation, deployment change information of the first model is obtained, the deployment change information is determined based on first record information, the first record information is generated by the computing node; In response to the second execution operation, the first model is updated according to the deployment change information, and the first feature representation sequence is determined through the computing node. Updating the first model includes updating the feature unit distribution and the network structure distribution. Media content is generated based on the first feature representation sequence.

[0006] Secondly, this paper also provides a media content generation device, which includes: The acquisition module is used to acquire the first request content and input the first request content into the first model, which is distributed across multiple computing nodes. The first response module is used to respond to the first execution operation by obtaining deployment change information of the first model, wherein the deployment change information is determined based on first record information, and the first record information is generated by the computing node. The second response module is used to respond to the second execution operation, update the first model according to the deployment change information, determine the first feature representation sequence through the computing node, and update the first model includes updating the feature unit distribution and network structure distribution; The content generation module is used to generate media content based on the first feature representation sequence.

[0007] Thirdly, this article also provides an electronic device, which includes: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the media content generation method as described herein.

[0008] Fourthly, this document also provides a storage medium containing computer-executable instructions that, when executed by a computer processor, are used to perform media content generation methods as described herein.

[0009] Fifthly, this document also provides a computer program product, including a computer program that, when executed by a processor, implements the media content generation method as described herein.

[0010] This paper provides a media content generation method, apparatus, and program product. The method includes obtaining first requested content and inputting the first requested content into a first model, the first model being distributed across multiple computing nodes; responding to a first execution operation, obtaining deployment change information of the first model, the deployment change information being determined based on first record information generated by the computing nodes; responding to a second execution operation, updating the first model based on the deployment change information, determining a first feature representation sequence through the computing nodes, and updating the first model including updating the feature unit distribution and network structure distribution; and generating media content based on the first feature representation sequence. This technical solution, combining a first model distributed across multiple computing nodes, achieves distributed media content generation. Specifically, the first model can perform content generation processing based on the first requested content. During the processing of the first model, after triggering the first execution operation, it can respond to the first execution operation. Thus, the deployment change information of the first model is determined through the first record information recorded by the computing nodes during the processing, and the feature unit distribution and network structure distribution of the first model can be adjusted based on the deployment change information. This ensures that the feature units in subsequent processing are located on the same computing node as the first model participating in feature unit processing, and that the load on each computing node is balanced. This effectively reduces the communication overhead in feature processing of computing nodes, and also effectively improves the processing speed of feature processing, thereby improving the efficiency of media content generation and ensuring the generation effect of media content. Attached Figure Description

[0011] The above and other features, advantages, and aspects of this technology will become more apparent when viewed in conjunction with the accompanying drawings and the following specific implementations. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0012] Figure 1 The paper provides a schematic diagram illustrating the application scenarios of the media content generation method presented in this paper. Figure 2 This is a flowchart illustrating a media content generation method presented in this paper. Figure 3 This is a schematic diagram illustrating the iterative feature processing flow of the media content generation method presented in this paper. Figure 4 This is a flowchart illustrating the process of determining the first record information in the media content generation method provided in this paper. Figure 5 A flowchart illustrating another media content generation method presented in this article; Figure 6 This is a schematic diagram illustrating the implementation of the media content generation method presented in this paper. Figure 7 This is a schematic diagram of the media content generation device provided in this paper. Figure 8 This is a schematic diagram of the electronic device used to implement the media content generation method provided in this paper. Detailed Implementation

[0013] The implementation of this technology will now be described in more detail with reference to the accompanying drawings. While some implementations of the technology are shown in the drawings, it should be understood that the technology can be implemented in various forms and should not be construed as limited to the implementations described herein. Rather, these implementations are provided to provide a more thorough and complete understanding of the technology. It should be understood that the accompanying drawings and implementations are for illustrative purposes only and are not intended to limit the scope of protection of the technology.

[0014] It should be understood that the steps described in the implementation of this technique may be performed in different orders and / or in parallel. Furthermore, the implementation of this technique may include additional steps and / or omit the steps shown. The scope of this technique is not limited in this respect.

[0015] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one implementation" means "at least one implementation"; the term "another implementation" means "at least one additional implementation"; the term "some implementations" means "at least some implementations". Definitions of other terms will be given in the following description.

[0016] It should be noted that the concepts of "first" and "second" mentioned in this technology are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0017] It should be noted that the terms "one" and "more" used in this technology are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] The names of the messages or information exchanged between the various devices in this article are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0019] It is understandable that before using the technical solutions disclosed in this article, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the technology and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0020] For example, upon receiving a user's active request, a prompt is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to autonomously choose whether to provide personal information to the software or hardware, such as electronic devices, applications, servers, or storage media, that are performing the operation of the technical solution, based on the prompt data.

[0021] As an optional but non-limiting implementation, in response to a user's active request, sending prompt data to the user can be done via a pop-up window, where the prompt data can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0022] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this technology. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this technology.

[0023] It is understood that the data involved in this article (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0024] For example, Figure 1 The paper provides schematic diagrams illustrating application scenarios for the media content generation method presented in this paper. In practical applications, the media content generation method provided by this technology can be used to... Figure 1 The interactive information present in the application scenario shown is processed.

[0025] Among them, such as Figure 1 As shown, terminal 102 communicates with server 104 via a communication network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on the cloud or other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc., which can generate media content generation requests by triggering functional items in the displayed interface and send them to server 104. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers. Server 104 can execute the media content generation method provided herein, and the generated media content can be fed back to terminal 102 for display.

[0026] Figure 2This is a flowchart illustrating the media content generation method presented in this paper. This method is applicable to scenarios involving media content generation, particularly text-to-image, image-to-image, text-to-video, and image-to-video applications. The method can be executed by a media content generation device, which can be implemented in software and / or hardware, optionally through an electronic device such as a mobile terminal, computer, or server. Figure 2 As shown, the method in this paper may specifically include: S201. Obtain the first request content and input the first request content into the first model, wherein the first model is distributed across multiple computing nodes.

[0027] In one scenario, the entity executing the method provided herein can be considered a terminal containing multiple computing nodes. These computing nodes are preferably graphics processors configured on the terminal; that is, the terminal can be considered to have multiple image processors. The terminal can be a mobile terminal, a computer, or a server, etc., and a first model is deployed on the terminal, with the first model of the first model deployed on different computing nodes.

[0028] The first model can be considered a network model for realizing media content generation tasks. This network model can be considered a combination of a hybrid expert model and a diffusion model, possessing long sequence processing capabilities. Each computing node on the terminal can carry a portion of the first model, and all computing nodes can simultaneously support the media content generation of the first model.

[0029] In one scenario, the terminal executes the media content generation method upon receiving a media content generation request. The execution of the media content generation method is initiated by responding to this request. The media content generation request can be triggered directly on the terminal acting as the executor, or it can be triggered by another terminal with a communication connection to the executor and then sent to the executor. For example, the triggering method for the media content generation request could be clicking a generation control, submitting a text command, or detecting a function call request. For example, this media content generation request can be understood as an interactive request for generating visual content, which may include image content and video content.

[0030] In one scenario, the media content generation request may include a description of the media content generation requirements. The request data can be extracted by parsing the request to obtain the first request content carrying the media content generation requirements. This first request content may include at least one of the following: a text description, reference media content (such as reference images or video clips), and generation parameters (such as screen resolution, frame rate, and style type). For example, the text description may describe the media content generation requirements, specifying whether an image or video is desired, and the desired generated content or style. The first request content may also include a reference image and a text description; for instance, the text description may be supplementary text to the reference image.

[0031] In one scenario, the first model needs to generate media content by repeatedly executing multiple rounds of processing logic. Each round of processing logic optimizes the detailed features involved in media content generation based on the output of the previous round, thereby gradually approximating the feature information of the final required media content. The network structure of the first model in this paper is distributed across different computing nodes. Therefore, each round of processing logic in the first model can be considered as the computing nodes providing the computing carrier to support the processing of the first model. Each round of processing logic in the first model can be regarded as a round of feature processing on the first requested content. Specifically, after executing a round of processing logic, a feature representation sequence associated with the first requested content is generated. The feature representation sequence generated in the previous round can be used as the input of the first model in the next round, further used for the generation of feature representation sequences. Compared with the feature representation sequence generated in the previous round, the feature representation sequence generated in the next round contains more detailed features involved in media content generation.

[0032] It should be noted that this paper firstly uses a general processing logic for media content generation to generate a sequence of feature representations in the first few iterations after the first model begins processing. The termination condition for executing this general processing logic can be the fulfillment of the triggering condition for the first execution operation. In one implementation, this triggering condition can be that the number of iterations of feature processing by the first model reaches a first threshold. Therefore, it can be assumed that when the number of iterations is less than the first threshold, the general processing logic is used in each round of feature processing, thereby generating a sequence of feature representations related to media content generation in each round. For example, the first threshold can be set as a percentage of the total number of iterations involved in the processing of the first model; that is, the first threshold can be the product of the total number of iterations and this percentage.

[0033] In one implementation, the execution logic of the first model performing feature processing iteratively based on the first request content can be described as follows: the obtained first request content is formatted to generate the first feature representation sequence that can be used as input data for the first model. The first model can divide the feature units in the first generated feature representation sequence into each computing node according to the existing feature splitting logic. Each computing node is equivalent to obtaining a part of the feature units in the feature representation sequence.

[0034] As described above, the first model on each computing node can synchronously start processing tasks. Specifically, the first model on the computing node can perform attention processing on the corresponding obtained feature units, then determine a matching expert network for the attention feature units after attention processing, and distribute the attention feature units to the matching expert networks. The expert networks then perform feature transformation processing on the attention feature units to obtain the transformed feature units. Thus, the expert networks can be deployed with other computing nodes. Therefore, the transformed feature units output by each expert network need to be aggregated to the computing node where the feature unit is located to form the feature units after feature processing. The first model can determine a new feature representation sequence based on the processed feature units, and then the first model on each computing node starts a new round of processing based on the feature representation sequence according to the above processing logic.

[0035] In media content generation, the feature representation sequence can be considered a summary of feature units. A feature unit can be considered the smallest unit representing the features of media content, corresponding to the features of an image block in the media content. Furthermore, the first model can be considered to contain at least an input layer, a feature processing layer, and an output layer. The feature processing layer may contain an attention module and an expert network module. The attention module can be used for attention processing of feature units, and the expert network module can be used for feature transformation processing of feature units. As described above, before responding to the first execution operation, the first model on each computing node will continuously perform feature processing iteratively according to the above description.

[0036] It should be noted that, under normal circumstances, once the first model is deployed on different computing nodes, it remains fixed. In each iteration, the feature units of the first model are also allocated to different computing nodes for feature processing according to a fixed feature allocation logic. However, this deployment and allocation method leads to high communication overhead for the first model, consuming computing resources and causing low computing speed due to load imbalance among computing nodes. The high communication overhead is because the feature processing stage of the first model requires determining the expert network for the feature units. Since the expert network is generally deployed on different computing nodes, this results in communication time consumption for feature distribution and result aggregation. This communication time consumption occurs in each iteration, and as the number of iterations accumulates, the communication time consumption increases, resulting in a significant consumption of computing resources due to communication time consumption.

[0037] Meanwhile, the existing deployment and allocation methods also suffer from load imbalance. Some computing nodes are allocated a large number of first models or feature units of the first model. As a result, the computing nodes will consume more processing time in supporting the operation of the first model. The aggregation of feature processing results on each computing node can only be started after the last computing node has finished processing. The time taken for a single iteration is lengthened, which will eventually lead to a significant decrease in the overall media content generation speed.

[0038] Based on the above description, this paper proposes to determine deployment change information in the execution of the media content generation method, and redeploy the distribution of the first model on the computing nodes according to the deployment change information. This distribution includes the deployment from feature units to computing nodes and the deployment of the first model to computing nodes. This distribution adjustment can solve the problems of high communication overhead and slow computing speed. It is equivalent to optimizing the existing static deployment of the first model into a dynamic deployment. However, it should be noted that if the deployment strategy is dynamically determined before each iteration and the first model is redeployed according to the deployment strategy determined in each iteration, the multiple determinations of the deployment strategy will also bring communication overhead, and the above problems cannot be solved better.

[0039] To better address the aforementioned issues, this paper fully considers the processing characteristics of the first model in feature processing. It finds that after the initial few rounds of feature processing, subsequent iterations show that the feature units remain largely consistent with the matched expert network, unaffected by iterations or changes in the size of the first model or the total number of iterations. Based on this, this paper proposes an optimized approach: after a certain number of rounds of feature processing, the determination of deployment change information is initiated, and this information is reused in subsequent rounds of feature processing.

[0040] It is understood that this paper can determine deployment change information after responding to the first execution operation. The triggering condition for this first execution operation can be that the number of iterations of feature processing performed by the first model reaches a first threshold. This is equivalent to the first model on the computing node having undergone multiple rounds of feature processing after the first threshold, ensuring that the feature units are consistent with the matched expert network, thereby initiating the determination of deployment change information.

[0041] S202. In response to the first execution operation, obtain the deployment change information of the first model, wherein the deployment change information is determined based on the first record information, and the first record information is generated by the computing node.

[0042] In one scenario, this step can be used to respond to the first execution operation after it has been triggered, thereby triggering the execution logic for determining deployment change information. This can be viewed as determining deployment change information for the first model, and this determination can be undertaken by one of multiple computing nodes. Therefore, this step can be considered to be executed by one of the computing nodes.

[0043] In one scenario, the deployment change information of the first model can be determined based on the first record information, which can be considered to be generated during the feature processing of the first model on the computing node. Specifically, it can be a record of the processing time generated by the computing node in each iteration, or a record of the allocation information determined in the feature processing. The allocation information can be considered to include the association information of the feature units and the matching expert network.

[0044] It is important to understand that the first model generates allocation information in each round of feature processing. Therefore, the first record information can be considered to correspond to the formation of each round of feature processing. This first record information essentially includes the processing time of the computing node in processing the feature units on the first model, as well as the allocation information between the feature units formed on the first model and the expert network. Based on this, this paper can specifically define determining the deployment change information for the first model as determining the feature deployment change information and the structural deployment change information of the first model respectively.

[0045] As one implementation method, the process of determining deployment change information for the first model can be described as follows: Based on the first record information of the first model, determine the feature deployment change information used for feature unit partitioning, and determine the structural deployment change information used for expert network deployment. Specifically, the determination of feature deployment change information can be achieved by determining the feature unit similarity matrix based on the allocation information in the first record information. This allocation information represents the association between feature units and the expert network. Then, feature units are clustered based on the feature unit similarity matrix, with the number of cluster sets being the same as the number of computing nodes. Thus, one computing node can be bound to one cluster set. Finally, considering load balancing constraints, the feature units in the cluster sets are adjusted. The feature units included in the adjusted cluster sets can be considered the most suitable feature units for partitioning onto computing nodes. The association information between each cluster set and the corresponding computing node constitutes the sub-feature deployment change information of the first model.

[0046] The determination of structural deployment change information can be specifically based on the aforementioned characteristic deployment change information combined with the first record information. Specifically, the characteristic deployment change information reveals the feature units contained in the computing node. Then, by combining this with the allocation information in the first record information, the demand degree of the computing node relative to each expert network on the first model can be determined. Subsequently, a pre-constructed first function can be obtained, which can be used to indicate the operating resources of the first model. These operating resources can include communication resources in the communication dimension and computing resources in the computation dimension. Finally, the expert network partitioning strategy can be determined based on the demand degree of the computing node relative to each expert network. This expert network partitioning strategy can provide an adapted expert network for each computing node, and the determination of the expert network partitioning strategy can use the minimum function value of the first function as a constraint.

[0047] This step can be described above to determine the feature deployment change information and the structural deployment change information of the first model. By summarizing the feature deployment change information and the structural deployment change information, the deployment change information of the first model can be obtained.

[0048] It is known that deployment change information ensures that feature units allocated on computing nodes can be assigned to the first model (expert network) deployed on that computing node for feature processing. This significantly reduces the communication overhead of distributing feature units to expert networks on other computing nodes, and also greatly reduces the communication overhead of merging the processing results from expert networks on other computing nodes back to this computing node. Furthermore, maintaining load balancing between expert networks and feature units on the computing node effectively improves computational speed.

[0049] S203. In response to the second execution operation, update the first model according to the deployment change information, and determine the first feature representation sequence through the computing node.

[0050] Based on the above description, it can be seen that after the deployment change information is determined, for each subsequent iteration with an iteration count greater than the first threshold, the deployment change information can be reused to update the distribution from the first model to the computing node. The iteration count being greater than the first threshold and the iteration not being completed can be used as the triggering condition for the second execution operation. Thus, for each iteration with an iteration count greater than the first threshold, the second execution operation can be triggered before the iteration ends. This step can repeatedly respond to the triggered second execution operation.

[0051] This step, following the response to the second execution operation, can be used to repartition the feature units of the first model to computing nodes and redeploy the expert network to computing nodes. The computing nodes can then support the first model in feature processing based on the repartitioned feature units and the redeployed first model, until the first model reaches the iteration termination condition. The redeployed expert network can be located in the feature processing layer of the first model.

[0052] In one scenario, updating the first model may include updating the feature unit distribution and the network structure distribution. Updating the feature unit distribution may involve updating the partitioning of feature units on computing nodes, while updating the network structure distribution may involve updating the deployment of the expert network included in the first model on computing nodes.

[0053] In one scenario, the process of updating the first model based on the deployment change information and determining the first feature representation sequence through the computing node can be described as follows: First, the feature units assigned to the computing node and the expert network deployed on the computing node are determined based on the deployment change information. Then, the computing node can perform feature processing on the assigned feature units relative to the first model based on the deployed expert network. The above logic can ensure that the first model completes one round of feature processing, and the feature representation sequence output by the first model can be used as the feature representation sequence formed by feature processing in this round of iteration.

[0054] Using the aforementioned deployment change information allows computing nodes to save more communication overhead in feature unit distribution and feature processing merging in the implementation of feature processing for the first model. Therefore, the first model can save more communication overhead by performing multiple rounds of feature processing, which in turn can greatly reduce communication overhead in the entire iterative feature processing. At the same time, distributing and updating the first model according to the deployment change information can also ensure that the computing nodes are in a balanced state, which also greatly improves the computing speed.

[0055] S204. Generate media content based on the first feature representation sequence.

[0056] In one scenario, this step can be considered to be executed when the first model has finished iterative feature processing, thus allowing the logic for generating media content to be executed through this step.

[0057] It is known that after completing the feature processing iteration, the first model on each computing node can output a first feature representation sequence. By merging each first feature representation sequence, a total feature representation sequence for media content generation can be obtained. Then, this total feature representation sequence can be input into the output layer of the first model, where the output layer decodes the total feature representation sequence and outputs the media content of the required generation type, which is the media content that matches the first requested content.

[0058] The media content can be generated as either an image or a video. The output layer can decode the total feature representation sequence based on the desired generation type of the media content. For example, when the generation type is an image, noise in the total feature representation sequence can be gradually removed by a noise sampler in the first model to convert the high-dimensional feature tensor into an image pixel tensor. Finally, the image details can be optimized by a super-resolution module to obtain a higher-resolution image as the media content. Alternatively, when the generation type is a video, the total feature representation sequence can be first decomposed into frames by time step to obtain the feature representation sequence of each frame. Then, the feature representation sequence of each frame can be decoded to generate an image. Finally, the image frames can be stitched together in chronological order, and / or an audio track can be added (if the first requested content includes audio requirements) to generate a complete video file as the media content.

[0059] The above technical solution combines a first model distributed across multiple computing nodes to achieve distributed media content generation. Specifically, the first model generates content based on a first requested content. During the processing of the first model, a first execution operation is triggered, and the model responds to this operation. By using the first record information recorded by the computing nodes during the processing, the deployment change information of the first model is determined. This deployment change information allows for adjustments to the distribution of feature units and the network structure of the first model, ensuring that subsequent feature units reside on the same computing node as the first model involved in feature unit processing, and maintaining a balanced load across all computing nodes. This effectively reduces communication overhead in feature processing on computing nodes, improves feature processing speed, and ultimately enhances media content generation efficiency and quality.

[0060] As described above, before the triggering condition of the first execution operation is met, a processing logic is needed to iteratively implement feature processing. As an optional technical solution provided above, this paper can provide a specific implementation of the first model iteratively performing feature processing based on the first request content. For example, Figure 3 This is a schematic diagram illustrating the iterative feature processing flow in the media content generation method presented in this paper, as follows: Figure 3 As shown, the steps may include: S301. Using the computing node, the input layer of the first model is used to perform feature parsing on the first request content to obtain a second feature representation sequence.

[0061] In one scenario, this step can be viewed as being performed by the input layer in the first model. This input layer can be deployed on one or more computing nodes of multiple computing nodes. Thus, the computing node with the input layer can perform feature parsing on the input first request content and convert the first request content into a feature tensor that can perform subsequent steps. This feature tensor can be a feature representation sequence containing multiple feature units of the smallest representation unit. This paper may also refer to this feature representation sequence as the second feature representation sequence.

[0062] S302. Using the feature processing layer of the first model, the second feature representation sequence is processed to obtain the third feature representation sequence.

[0063] In one scenario, the first model can be considered to include a feature processing layer. The second feature representation sequence obtained through the above steps can enter the feature processing layer, which can include two network modules: one can be an attention module and the other can be an expert network module.

[0064] In one scenario, it can be assumed that the feature processing layer is preferably distributed across multiple computing nodes. Each computing node can be considered to have deployed a portion of the feature processing layer in the first model. To ensure the normal operation of the first model, each computing node can run its respective feature processing layer in parallel.

[0065] In one scenario, the processing logic of a single round of feature processing performed by the feature processing layer of a computing node can be concretized as the feature processing layer on the computing node performing feature processing on the second feature representation sequence received by the input layer. Specifically, the computing node can obtain a subset of feature units from the second feature representation sequence, and these feature units can be allocated to the computing node using existing feature unit partitioning logic.

[0066] In one scenario, each computing node that has obtained the divided feature units can perform feature processing on the feature units it possesses. After completing the processing of each first model, each computing node can output a new feature representation sequence, which can be denoted as the third feature representation sequence.

[0067] In one scenario, processing the second feature representation sequence to obtain the third feature representation sequence may include: using the attention module to perform attention processing on the second feature representation sequence to obtain an attention feature representation sequence; using the expert network module to assign the first feature unit to the first expert network for feature transformation according to the first allocation information, obtaining the feature transformation result output by the first expert network, and obtaining the third feature representation sequence based on the feature transformation result, wherein the attention feature representation sequence contains the first feature unit, and the first allocation information records the allocation information of the first feature unit to the first expert network.

[0068] In this context, it can be assumed that the feature processing layers on each computing node have the same processing logic. In the first round of processing of the feature processing layer, its second feature representation sequence can come from the feature representation sequence output by the input layer. The feature processing layer can first perform attention processing on the second feature representation sequence through the attention module, thereby obtaining the attention feature representation sequence output by the attention module.

[0069] As described above, the attention feature representation sequence can be used as input to the expert network module included in the feature processing layer. The expert network module first determines the allocation information for the attention feature representation sequence. This allocation information is denoted as the first allocation information, which includes the mapping information of assigning a suitable expert network to the feature unit. That is, it includes the mapping information of the expert network adapted to each feature unit. The feature unit can be denoted as the first feature unit included in the attention feature representation sequence, and the adapted expert network can be denoted as the first expert network. Different first expert networks may be deployed on different computing nodes.

[0070] It is important to understand that, based on the existing static partitioning and deployment logic for feature unit partitioning and expert network deployment, it cannot be guaranteed that the first feature unit and the adapted first expert network will all be deployed on the same computing node. Therefore, in the implementation of allocating the first feature unit to all adapted first expert networks using the first allocation information, communication overhead is required for first expert networks deployed on computing nodes other than the first computing node to communicate with other computing nodes.

[0071] In one scenario, after the first expert network is assigned to the first feature unit, it can perform feature transformation processing on the first feature unit. The feature transformation results output by the first expert network will also be fed back to the computing node. That is, the computing node can obtain the feature transformation results output by each first expert network. The computing node needs to perform weighted fusion on the feature transformation results of each first feature unit, and can form the third feature representation sequence output by the feature processing layer based on the obtained weighted fusion results.

[0072] The above description illustrates the implementation of feature processing on the second feature representation sequence by the feature processing layer on the computing node. In the iterative feature processing, each round of feature processing can be performed in the manner described above, and the third feature representation sequence output from the previous round can be used as the input to the second feature representation sequence for the next round of feature processing. This process continues until the first execution operation is responded to. This implementation ensures the normal progress of feature processing in the first model and provides fundamental data support for determining the first record information.

[0073] S303. In response to the third execution operation, the third feature representation sequence is used as the second feature representation sequence, and S302 is re-executed.

[0074] The above steps are equivalent to completing one round of feature processing through computing nodes, with each computing node obtaining a corresponding third feature representation sequence. This paper considers the triggering condition for the third execution operation to be met when the iteration termination condition is not met and the iteration count has not reached the first threshold. Therefore, this step can be used to respond to the third execution operation and merge the third feature representation sequences output by each computing node to form a new second feature representation sequence.

[0075] In one scenario, this step can initiate a new round of feature processing on the second feature representation sequence, specifically by returning to step S302 above and re-executing. It should be noted that before initiating a new round of feature processing on the computing nodes, the existing partitioning and deployment logic will be used to re-partition the second feature representation sequence into feature units and deploy the network structure on the first model. The feature units partitioned on each computing node can then be considered the new second feature units.

[0076] The above technical solution is equivalent to implementing iterative feature processing of the first model through computing nodes according to a feature processing logic. After the first model starts running, it can ensure the normal progress of feature processing in the first few iterations through the above technical solution. However, it is also known that the first model will incur communication overhead on each computing node in each round of feature processing. When the feature units are unevenly distributed on the computing nodes and / or the expert network is unevenly deployed, the first model waits for the last computing node to complete the feature processing. This uneven distribution or deployment will also affect the feature processing speed. Therefore, if the feature processing in media content generation is continuously promoted according to the above feature processing logic, the problems of large communication overhead and low computational efficiency will arise.

[0077] Simultaneously, during the processing according to the aforementioned feature processing logic, first record information can be formed, thereby providing data information support for the subsequent determination of deployment change information. As an optional technical solution provided above, this paper can also determine the first record information through computing nodes during the iterative feature processing of the first model based on the first request content. Figure 4 This is a flowchart illustrating the process of determining the first record information in the media content generation method provided in this paper, as shown below. Figure 4 As shown, the provided media content generation method may also include: S401. The processing time for the first model is obtained through the computing node, wherein the processing time is related to the processing of the first request content using the first model.

[0078] In one scenario, this article can be considered... Figure 4 The determination logic shown can be performed during the feature processing of each iteration. Furthermore, the determination logic for the first record information is the same in each iteration of feature processing.

[0079] In one scenario, this step can obtain the processing time consumed by the computing node when the first model performs feature processing. This processing time can include the communication time and computation time consumed by the computing node during the runtime of the first model. Specifically, this processing time can be related to the feature processing performed by the first model on the first requested content. The communication time can be considered as the communication time spent by the computing node when the first model processes the feature units, and the computation time can be considered as the computation time spent by the computing node when the first model processes the feature units. This computation time can include the computation time for attention processing and the computation time for feature transformation processing performed by the expert network.

[0080] S402. In response to the first acquisition operation, acquire second allocation information, the second allocation information being related to the allocation of feature units on the first model to the second expert network, the second expert network being included in the first model.

[0081] In one scenario, the above steps can be used to obtain the processing time of the computation node in each iteration. However, it is necessary to obtain the second allocation information during the feature processing performed by the first model in this step only when the acquisition condition is met. This acquisition condition can be considered the generation condition of the first acquisition operation, which can be that the number of iterations has reached a second threshold, and this second threshold is less than the first threshold. It can be any value less than the first threshold, but generally needs to be greater than half of the first threshold. For example, when the first threshold is 10, the second threshold can be 6.

[0082] In one scenario, the second allocation information can be considered as the allocation information required for feature transformation processing by the expert network module in the first model on the computing node. This second allocation information can also be determined by the expert network module, specifically indicating which allocable expert networks are associated with the feature units to be processed by the expert network model. This allocable expert network can be denoted as the second expert network, which is included in the first model and belongs to the expert network module.

[0083] S403. Record the processing time and the second allocation information to form the first recording information.

[0084] In one scenario, this step can be used to record the second allocation information and the processing time of the computing nodes corresponding to the first model, thereby forming the first record information.

[0085] In this optional technical solution, the determination of the first record information can serve as data support for the determination of subsequent deployment change information, providing a quantifiable decision-making basis for deployment change information and ensuring data support in both communication and load balancing dimensions. This first record information also indirectly verifies and quantifies the characteristic that the feature units of the first model and the expert network maintain consistency with each iteration. This supports the one-time determination and multiple reuse of deployment change information.

[0086] As can be seen, the above technical description clearly states that when responding to the first execution operation, the logic for determining the deployment change information is initiated. At the same time, the determination of the deployment change information, after satisfying the generation of the second execution operation, is reused in the response of the second execution operation to effectively reduce the communication overhead in the entire iterative feature processing and improve the computation speed of feature processing.

[0087] As an optional technical solution provided in this article, Figure 5This is a flowchart illustrating another media content generation method provided in this paper. This optional technical solution can be combined with other implementations. For the same or related parts, descriptions of other implementations can be used, and will not be repeated here. Figure 5 As shown, the specific steps of the media content generation method in this paper may include: S501. Obtain the first request content and input the first request content into the first model, wherein the first model is distributed across multiple computing nodes.

[0088] S502. In response to the first execution operation, obtain the second allocation information included in the first record information, and determine the feature deployment change information of the first model based on the second allocation information.

[0089] In one scenario, the first record information corresponds to the first model. The first record information includes the cumulative processing time of the computing nodes in each first model when performing feature processing according to the above feature processing logic, as well as the fusion of the allocation information used by the expert network module in each iteration round when performing feature transformation processing on the first model.

[0090] In one scenario, this step first uses the second allocation information in the first record information to determine the feature deployment change information for the first model. Here, the feature deployment change information can be understood as the feature unit partitioning strategy upon which the feature units are partitioned in the first model.

[0091] In one scenario, the determination of feature deployment change information can be based on the second allocation information in the first record information to determine the expert network adapted to each feature unit. The similarity between pairs of feature units can be determined based on the expert network adapted to each feature unit. Then, the feature units can be clustered based on the similarity between pairs of feature units, specifically obtaining multiple feature cluster sets. The number of obtained feature cluster sets is the same as the number of computing nodes, which is equivalent to binding a feature cluster set to each computing node. Then, the number of feature units in the feature cluster set bound to each computing node can be adjusted to determine that the feature units associated with each computing node are in a load-balanced state. Thus, the feature units in the feature cluster set finally associated with the computing node can be regarded as the feature units most suitable for being assigned to the computing node. The feature cluster sets of each computing node on the first model constitute the feature deployment change information of the first model.

[0092] It is important to know that feature clusters can use the location identifier of the feature unit to represent the feature unit, which can be the subscript value of the feature unit in the feature representation sequence.

[0093] In one scenario, determining the feature deployment change information of the first model based on the second allocation information may include: obtaining the second allocation information from the first record information; determining the feature unit similarity matrix of the first model based on the second allocation information; determining the second feature unit associated with the computing node based on the feature unit similarity matrix, wherein the second feature unit is on the first model; and generating feature deployment change information based on the computing node and the second feature unit.

[0094] In one scenario, determining the feature unit similarity matrix can be described as follows: Based on the second allocation information in the first record information, obtain the expert networks associated with each feature unit. Then, for any pair of feature units, determine the number of overlapping expert networks in which both feature units in the pair have expert networks. The ratio of this number of overlapping expert networks to the total number of expert networks associated with the feature unit can be used as the feature unit similarity between the two feature units in the pair. This paper can use this logical description to determine the feature unit similarity between any two feature units, thereby constructing a feature unit similarity matrix based on the similarity of each feature unit. The number of rows and columns of this feature unit similarity matrix are both the number of feature units possessed by the first model.

[0095] In one scenario, feature units can be clustered using the similarity between pairs of feature units represented in the feature unit similarity matrix. The clustering implementation can use maximizing the similarity of feature units within a cluster and minimizing the similarity of feature units between clusters as constraints. Furthermore, the number of clusters can be set to be the same as the number of computing nodes, thereby binding a feature cluster set to each computing node. Thus, the second feature unit associated with the computing node can be obtained through the feature cluster set. This first feature unit can be considered as a feature unit suitable for assignment to computing nodes in the first model.

[0096] In one scenario, one implementation of determining the second feature unit associated with the computing node based on the feature unit similarity matrix can be described as follows: Based on the feature unit similarity matrix, clustering is performed on the network feature units with the goal of maximizing intra-cluster feature unit similarity and minimizing inter-cluster feature unit similarity, where the network feature units are the feature units present in the first model; a first number of feature cluster sets is obtained, the first number being the same as the total number of computing nodes; the feature cluster sets are bound to the computing nodes according to the node sequence number and cluster set sequence number to form the first feature unit set of the computing node; the feature units included in the first feature unit set are adjusted according to the node load balancing condition, and the feature units in the adjusted first feature unit set are determined as the first feature unit.

[0097] Specifically, setting the number of feature cluster sets to be the same as the total number of computing nodes ensures that each computing node is bound to a feature cluster set. The node load balancing condition can be a feature unit allocation threshold, which can be the ratio of the total number of feature units to the total number of computing nodes. The adjustment logic for the feature units contained in the first feature unit set can be described as follows: select the feature unit with the smallest feature unit similarity value to the feature units in the set, and assign that feature unit to the first feature unit set with the fewest feature units. This process is repeated until the number of feature units in each first feature unit set is consistent.

[0098] The above steps are based on the feature unit similarity matrix, and cluster the network feature units around the constraints of maximizing intra-cluster similarity and minimizing inter-cluster similarity. This results in a feature cluster set consistent with the number of computing nodes, which is then bound by sequence number to form the first feature unit set. The feature units within the set are then optimized by combining node load balancing conditions. This approach not only aggregates highly similar feature units through accurate clustering, significantly reducing cross-node communication overhead, but also ensures that the number of feature units carried by each computing node is consistent with the load through load balancing, avoiding single-node overload becoming an iteration bottleneck. At the same time, the sequence number binding method simplifies the deployment logic and improves adjustment efficiency. Ultimately, while ensuring the continuity and accuracy of feature unit processing, it significantly improves the iteration processing speed and overall media content generation efficiency.

[0099] In one scenario, after identifying the associated first feature unit for each computing node, the association information representing the relationship between each computing node and its associated first feature unit can be determined as feature deployment change information.

[0100] The aforementioned determination of feature deployment change information enables refined and data-driven optimization from feature units to computing nodes in the first model. This allows for more balanced computing node load and significantly reduces cross-node communication overhead. Furthermore, the determination of this information relies on the quantitative support of the first record information, ensuring the traceability and adjustability of the feature deployment change information determination. Ultimately, without changing the network structure of the first model, this provides a guarantee for improving the processing efficiency of subsequent feature processing and the speed of media content generation.

[0101] S503. Based on the feature deployment change information and the second allocation information, determine the structural deployment change information of the first model.

[0102] In one scenario, this step can determine the structural deployment change information for the first model using the second allocation information and the aforementioned determined feature deployment change information. The structural deployment change information can be understood as the expert network deployment strategy required to deploy the expert network on the first model to the computing nodes.

[0103] In one scenario, determining structural deployment change information can begin by first identifying the feature units of the feature clusters corresponding to each computing node, combined with the second allocation information in the first record information, to determine which expert networks are suitable for deployment on the computing nodes. Based on the expert networks suitable for deployment on each computing node, the demand degree of the computing node relative to the expert networks can be statistically calculated. Then, based on the demand degree of the computing node relative to the expert networks, and using the minimum function value of the constructed first function as a constraint, this paper needs to solve for the expert network associated with the computing node that guarantees the minimum function value of the first function. This expert network is then considered as the final suitable expert network for deployment on the computing node, possessing characteristics that adapt to the feature units partitioned on the computing node.

[0104] In one implementation, a first function can be used to indicate the runtime resources of a first model. The first function may include a first sub-function that indicates the resource consumption related to the communication dimension, and may also include a second sub-function that indicates the resource consumption related to the computation dimension. The processing time of the computing nodes contained in the first record information can be used as a weighting coefficient in the constructed first function. The communication time in the processing time can be used as a weighting coefficient of the first sub-function, and the computation time can be used as a weighting coefficient of the second sub-function.

[0105] In one scenario, determining the structural deployment change information of the first model based on the feature deployment change information and the second allocation information may include: obtaining a first function, wherein the first function is related to the running resources of the first model; determining the expert network demand degree of the computing node relative to the first model based on the feature deployment change information and the second allocation information; determining a third expert network associated with the computing node based on the expert network demand degree, wherein the third expert network belongs to the first model and satisfies a first condition, wherein the first condition is that the function value of the first function is minimized; and generating structural deployment change information based on the computing node and the third expert network.

[0106] The process of determining the demand degree of a computing node relative to the expert network can be as follows: Using the second allocation information, determine the expert networks corresponding to each first feature unit associated with the computing node. Then, by analyzing the association between the first feature units and the expert networks, determine how many first feature units the expert network is compatible with. The number of compatible first feature units can be considered as the demand degree of the computing node relative to the expert network. The first function includes a first sub-function corresponding to the communication dimension, which can be a function that accumulates the communication cost of allocating feature units to the expert network. The first function also includes a second sub-function corresponding to the computation dimension, which can be a function that accumulates the computational cost involved in the computing node using the expert network to complete feature processing.

[0107] The aforementioned determination of structural deployment change information enables refined and data-driven optimization of the expert network to computing nodes in the first model. This allows for more balanced computing node load and a significant reduction in cross-node communication overhead. Furthermore, the determination of this information relies on the quantitative support of the first record information, ensuring the traceability and adjustability of the feature deployment change information determination. Ultimately, without altering the network structure of the first model, this provides a guarantee for improving the processing efficiency of subsequent feature processing and the speed of media content generation.

[0108] The above steps combine the first record information of the first model with the feature deployment change information, which can accurately determine the demand of computing nodes for each expert network. At the same time, combined with the first function, the expert network that minimizes the function value can be selected as the third expert network. This achieves accurate matching between the expert network and the needs of computing nodes, reducing cross-node communication overhead. Furthermore, the constraints of the first function ensure balanced resource consumption of each computing node, avoiding efficiency bottlenecks caused by unreasonable allocation of expert networks. Ultimately, without affecting the continuity of feature processing, it can further guarantee the stability and accuracy of feature processing efficiency.

[0109] S504. Determine deployment change information based on the feature deployment change information and structure deployment change information.

[0110] In one scenario, the feature deployment change information and structural deployment change information determined around the first model can be denoted as deployment change information.

[0111] S505. In response to the second execution operation, according to the feature deployment change information, the third feature unit is assigned to the associated computing node, wherein the third feature unit is on the first model.

[0112] It can be assumed that before starting a new round of feature processing, the computing nodes will perform initial feature unit partitioning and expert network deployment according to the existing partitioning and deployment logic. However, this initial partitioning and deployment suffers from problems such as high communication overhead and low computational efficiency.

[0113] As described above, the deployment change information determined using the above steps can be reused multiple times in subsequent feature processing iterations after a single determination. Compared to the original deployment information used for feature unit partitioning and expert networks, this deployment change information ensures that the feature units partitioned on the computing node are perfectly compatible with the expert network deployed on that computing node. This significantly reduces communication overhead and greatly improves the computational speed of feature processing after rearranging feature units and redeploying the expert network according to the deployment change information.

[0114] In one scenario, the condition for reusing the deployment change information may be in response to a second execution operation, which is triggered when the number of iterations of feature processing exceeds a first threshold and the iteration execution has not yet ended.

[0115] In one scenario, after responding to the second execution operation, the reuse of deployment change information in feature unit partitioning can be achieved through this step. Specifically, this step can be used to obtain feature units allocated to computing nodes according to the feature deployment change information in the deployment change information, and these feature units can be denoted as the third feature unit. This third feature unit exists in the first model, and it can exist in a feature representation sequence. This feature representation sequence can be the feature representation sequence output by the first model after processing through the feature processing logic given above, or it can be the feature representation sequence output by the first model after reusing the deployment change information and performing feature processing again.

[0116] S506. According to the structural deployment change information, deploy the fourth expert network to the associated computing nodes, wherein the fourth expert network belongs to the first model.

[0117] In one scenario, after responding to the second execution operation, the reuse of deployment change information in the expert network deployment can be achieved through this step. Specifically, this step can be used to obtain the expert network deployed to the computing node according to the structural deployment change information in the deployment change information, and this expert network can be referred to as the fourth expert network. This fourth expert network belongs to the first model, specifically to the expert network module of the feature processing layer on the first model.

[0118] The above steps enable precise adaptation of feature units and expert networks on the same computing node by deploying change information, reducing cross-node communication and load imbalance issues, and better ensuring the continuity and integrity of feature processing. At the same time, the optimization of the first model deployment can be completed without manual intervention, significantly improving the feature processing efficiency of the first model.

[0119] It can be understood that deploying change information provides information support for the partitioning of feature units involved in the first model on each computing node and the deployment of the expert network involved. After completing the partitioning of feature units to computing nodes and the deployment of expert networks to computing nodes using the above steps, the feature processing of the first model can be supported through the computing nodes. The following S507 and S508 give the specific implementation of feature processing of the first model.

[0120] S507. Through the computing node, attention processing is performed on the fourth feature representation sequence to obtain the fifth feature representation sequence. The fourth feature representation sequence is the input content of the first model. The fifth feature representation sequence contains fourth feature units, which are divided according to the deployment change information.

[0121] In one scenario, the fourth feature representation sequence is the input content of the first model. This input content can be the feature representation sequence output by the first model after processing the feature processing logic given above before the deployment change information is determined, or it can be the feature representation sequence output by the first model after reusing the deployment change information and performing a round of feature processing through S507 and S508 after the deployment change information is determined.

[0122] In one scenario, the first model may include an attention module and an expert network module. The computing node can first perform attention processing on the fourth feature representation sequence through the attention module of the first model, thereby obtaining the feature representation sequence output by the attention module.

[0123] In one scenario, the feature representation sequences output by the attention modules on each computing node can be integrated again. The integrated feature representation sequence can be used as the feature representation sequence output by the attention module on the first model. This technique can then re-divide the integrated feature representation sequence to the computing nodes according to deployment change information. The feature representation sequence obtained by the computing nodes after re-division can be denoted as the fifth feature representation sequence. This fifth feature representation sequence contains the fourth feature unit divided according to the deployment change information.

[0124] S508. The fourth feature unit is assigned to the fifth expert network for processing according to the second allocation information, the fifth feature unit output by the fifth expert network is obtained, and the first feature representation sequence is obtained according to the fifth feature unit, wherein the second allocation information is included in the first record information, and the fifth expert network is deployed according to the deployment change information.

[0125] In one scenario, the computing node can use the fifth feature representation sequence as input to the expert network module in the first model. Thus, the computing node can assign a suitable expert network to the fourth feature unit in the fifth feature representation sequence according to the second allocation information. This expert network can be referred to as the fifth expert network.

[0126] The fifth expert network can perform feature transformation on the fourth feature unit, thereby outputting the fifth feature unit. This fifth feature unit needs to be fed back to the computing nodes, which then form the first feature representation sequence based on it. In one scenario, the computing nodes can perform weighted fusion of the fifth feature units corresponding to the same fourth feature unit, and then summarize the weighted fusion results of each fourth feature unit to form the first feature representation sequence.

[0127] The second allocation information can be considered to be included in the first record information, used to record each feature unit and its corresponding expert network, specifically including the record of the fifth expert network adapted to the fourth feature unit. In this step, the fifth expert network can be considered to have been pre-deployed to the computing node according to the deployment change information. According to the determination description of the deployment change information, it can be known that the fifth expert network is basically assigned to the same computing node as the fourth feature unit. Through this division and deployment, the communication consumption of the fourth feature unit being allocated to the fifth expert network can be greatly reduced, as can the communication consumption of the fifth expert network feeding back the fifth feature unit to the computing node, thereby greatly reducing the resource consumption of the first model in feature processing.

[0128] It is known that as long as the first model has not finished the iterative execution of feature processing, the generation of the second execution operation can be repeatedly triggered. Thus, by responding to the second execution operation, the above-mentioned S505 to S508 can also be repeatedly executed. It is also known that when the first model performs a new round of feature processing through S505 to S508, the first feature representation sequences generated by each computing node after the previous round of feature processing will be merged into a new feature representation sequence as the input feature representation sequence of the first model to participate in the next round of feature processing.

[0129] It should be noted that the execution of S505 to S508 can be sequential (e.g., Figure 5 As shown, multiple steps can be executed simultaneously. For example, the feature unit repartitioning implementation in S505 can be performed in parallel with the attention processing implementation in S507. In this case, it can be considered that the first model mainly contains the fifth feature representation sequence, so its partitioning of the third feature unit to the computing node can be equivalent to partitioning the fourth feature unit in the fifth feature representation sequence. The expert network redeployment implementation in S506 can also be performed in parallel with the expert network processing implementation in S508. Since the implementation of steps S505 and S506 mainly generates communication overhead, it takes up almost no processing time. When S505 and S507 are executed in parallel, and S506 and S508 are executed in parallel, the resource consumption caused by updating the first model can be effectively saved.

[0130] S509. Generate media content based on the first feature representation sequence.

[0131] For example, after the first model finishes iterative feature processing, this step can be used to generate media content. The first feature representation sequence can be considered as the feature representation sequence output by the first model on the computing node after completing the final round of feature processing.

[0132] In one scenario, it is necessary to fuse the first feature representation sequences output by the first model on each computing node, and use the fused feature representation sequence as input information for the output layer of the first model to participate in the generation of media content.

[0133] The above process steps are equivalent to starting with the acquisition of the first requested content, firstly generating deployment change information for the first model based on the first record information through the first execution operation, then updating the feature unit distribution and network structure deployment on the first model according to the deployment change information through the second execution operation, and finally generating media content by generating the first feature representation sequence output by the first model after iterative feature processing. This process leverages the time consistency inherent in model iteration to achieve one-time determination and multi-round reuse of deployment change information, avoiding additional overhead caused by high-frequency adjustments during the first model's operation. Simultaneously, the determined deployment change information also ensures load balancing of computing nodes, effectively guaranteeing the continuity of feature processing and the accuracy of the generated results. This paper effectively improves processing efficiency and media content generation speed without changing the first model's network architecture, and also significantly enhances the media content generation effect.

[0134] For example, the specific implementation of the media content generation method provided in this paper can be described using an application scenario involving text-to-image generation: This could be a client-server interaction scenario where the client-side consumer inputs a text request for image generation on a corresponding input interface, and sends this request, encapsulated in a media content generation request, to the server by triggering a control. The server can be the execution entity of the provided media content generation method, first parsing the request text from the media content generation request as the initial request content. Simultaneously, the server can have multiple computing nodes.

[0135] The first model can be distributed across multiple computing nodes. The first request content can be used as input information. The computing nodes can then use the input layer of the first model to perform feature parsing on the first request content, obtaining a second feature representation sequence. The computing nodes can also use the feature processing layer of the first model to perform iterative feature processing based on the second feature representation sequence. During feature processing, the first record information of the first model is determined. After the number of iterations reaches a certain threshold, a computing node can determine deployment change information based on the first record information. In the implementation of the media content generation method, the deployment change information can be determined once and reused multiple times. In each subsequent round of feature processing, the deployment change information can be used to update the distribution of the first model on the computing nodes. Specifically, it can include updating the distribution of feature units and updating the distribution of network structure to ensure that the expert network adapted to the feature unit is as close as possible to the same computing node, and to ensure that the feature units and expert networks on each computing node are in a load-balanced state.

[0136] By updating the computation nodes of the feature units and the expert network distribution, feature transformation processing can be performed on the divided feature units in conjunction with the deployed expert network to generate the first feature representation sequence. The above logic can be reused in each round of feature processing until the feature processing iteration ends. Ultimately, this paper proposes that the first feature representation sequence output by each computing node in the last round can be used as the input information of the output layer of the first model. The output layer can generate media content by decoding the first feature representation sequence. This media content can be fed back to the client by the server and displayed on the client.

[0137] Similarly, the exemplary implementation of the media content generation method given above can also be achieved through... Figure 6 To provide an explanation, in which, Figure 6 This can serve as a schematic diagram of the media content generation method provided in this paper. It is described from the perspective of iterative feature processing using a first model. Specifically, the first model can be considered to be distributed across multiple computing nodes, such as... Figure 6 As shown, the first model 61 can start n rounds (the number of iterations at the end of the iteration) of iterative feature processing after receiving the first request content.

[0138] First, from round 1 to round t (t can be a small value relative to n, or alternatively, one-tenth of the value of n), feature processing can be performed by computing nodes according to the existing processing logic, and the feature processing process is recorded to form the first record information.

[0139] Secondly, in round t+1, it can be considered that the conditions for determining the deployment change information have been met. Therefore, while the computing nodes are performing feature processing according to the existing processing logic, one of the computing nodes can determine the deployment change information based on the first record information and obtain the deployment change information.

[0140] Then, the deployment change information can be passed to each round after round t+1. For example, from round t+2 to round n, the distribution of the first model on the computing nodes can be updated according to the deployment change information. This can include the distribution update of the feature units on the first model and the distribution update of the expert network on the first model. The computing nodes can then perform feature processing based on the feature units they possess and the redeployed expert network.

[0141] Because the distributed updates of the first model ensure that the feature units and their adapted expert networks are on the same computing node as much as possible, the resource consumption of the first model during runtime can be greatly reduced, the speed can be improved, the cost can be reduced, and the effect of media content generation can be guaranteed.

[0142] Figure 7 Here is a schematic diagram of the structure of the media content generation device provided in this article, such as... Figure 7 As shown, the device includes: Acquisition module 71 is used to acquire the first request content and input the first request content into the first model, wherein the first model is distributed across multiple computing nodes; The first response module 72 is used to respond to the first execution operation and obtain the deployment change information of the first model. The deployment change information is determined based on the first record information, which is generated by the computing node. The second response module 73 is used to respond to the second execution operation, update the first model according to the deployment change information, determine the first feature representation sequence through the computing node, and update the first model includes updating the feature unit distribution and the network structure distribution. The content generation module 74 is used to generate media content based on the first feature representation sequence.

[0143] The technical solution presented in this paper combines a first model distributed across multiple computing nodes to achieve distributed media content generation. Specifically, the first model generates content based on a first requested content. During the processing of the first model, a first execution operation is triggered, and the model responds to this operation. By using the first record information recorded by the computing nodes during the processing, the deployment change information of the first model is determined. This deployment change information allows for adjustments to the distribution of feature units and the network structure of the first model, ensuring that subsequent feature units reside on the same computing node as the first model involved in feature unit processing, and maintaining a balanced load across all computing nodes. This effectively reduces communication overhead in feature processing on computing nodes, improves processing speed, and ultimately enhances media content generation efficiency and quality.

[0144] Furthermore, the device may include a sequence generation module, which may specifically include: The parsing unit is used to perform feature parsing on the first request content through the computing node using the input layer of the first model to obtain a second feature representation sequence; The processing unit is used to process the second feature representation sequence using the feature processing layer of the first model to obtain the third feature representation sequence; An iterative unit is configured to, in response to a third execution operation, take the third feature representation sequence as the second feature representation sequence and return to re-execute the feature processing layer using the first model to process the second feature representation sequence and obtain the third feature representation sequence.

[0145] Furthermore, the feature processing layer includes an attention module and an expert network module. Specifically, the processing unit can be used to perform attention processing on the second feature representation sequence using the attention module to obtain an attention feature representation sequence; and to use the expert network module to assign the first feature unit to the first expert network for feature transformation according to the first allocation information, obtain the feature transformation result output by the first expert network, and obtain the third feature representation sequence according to the feature transformation result. The attention feature representation sequence contains the first feature unit, and the first allocation information records the allocation information of the first feature unit to the first expert network.

[0146] Furthermore, the device may also include an information determination module, configured to obtain, through the computing node, the processing time for the first model, the processing time being related to the processing of the first request content using the first model; in response to a first acquisition operation, acquire second allocation information, the second allocation information being related to the allocation of feature units on the first model to a second expert network, the second expert network being included in the first model; and record the processing time and the second allocation information to form the first record information.

[0147] Furthermore, the device may also include an information determination module, which may specifically include: The first determining unit is configured to acquire the second allocation information included in the first recording information, and determine the feature deployment change information of the first model based on the second allocation information; The second determining unit is used to determine the structural deployment change information of the first model based on the feature deployment change information and the second allocation information; The third determining unit is used to determine the deployment change information based on the feature deployment change information and the structure deployment change information.

[0148] Furthermore, the first determining unit can be specifically used to obtain second allocation information from the first record information, determine the feature unit similarity matrix of the first model according to the second allocation information; determine the second feature unit associated with the computing node according to the feature unit similarity matrix, wherein the first feature unit is on the first model; and generate feature deployment change information according to the computing node and the second feature unit.

[0149] Further, the second determining unit can be specifically used to obtain a first function, which is related to the running resources of the first model; determine the expert network demand degree of the computing node relative to the first model based on the feature deployment change information and the second allocation information; determine the third expert network associated with the computing node based on the expert network demand degree, the third expert network belonging to the first model, the third expert network satisfying a first condition, the first condition being that the function value of the first function is minimized; and generate structural deployment change information based on the computing node and the third expert network.

[0150] Furthermore, the second response module 73 can be specifically used to: divide a third feature unit to an associated computing node according to the feature deployment change information, wherein the third feature unit is on the first model; and deploy a fourth expert network to an associated computing node according to the structure deployment change information, wherein the fourth expert network belongs to the first model; wherein the feature deployment change information and the structure deployment change information are included in the deployment change information.

[0151] Furthermore, the second response module 73 can also be used to: perform attention processing on the fourth feature representation sequence through the computing node to obtain a fifth feature representation sequence, wherein the fourth feature representation sequence is the input content of the first model, the fifth feature representation sequence contains fourth feature units, and the fourth feature units are divided according to the deployment change information; allocate the fourth feature units to the fifth expert network for processing according to the second allocation information, obtain the fifth feature units output by the fifth expert network, and obtain the first feature representation sequence according to the fifth feature units, wherein the second allocation information is included in the first record information, and the fifth expert network is deployed according to the deployment change information.

[0152] Furthermore, the triggering condition for the first execution operation is that the number of iterations of feature processing by the first model reaches a first threshold; the triggering condition for the second execution operation is that the number of iterations is greater than the first threshold and the iteration has not ended; the triggering condition for the third execution operation is that the number of iterations is less than the first threshold.

[0153] The media content generation device provided in this paper can execute the media content generation method provided in any implementation of this technology, and has the corresponding functional modules and beneficial effects for executing the media content generation method.

[0154] It is worth noting that the various units and modules included in the above-mentioned device are divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of this document.

[0155] The following is for reference. Figure 8 This document illustrates a schematic diagram of an electronic device (e.g., a terminal device or server) 800 suitable for implementing a media content generation method. The terminal device referred to herein may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, personal digital assistants, tablets, portable multimedia players, in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital televisions and desktop computers. Figure 8 The electronic device shown is merely an example and should not impose any limitations on the functionality and scope of this article.

[0156] like Figure 8As shown, the electronic device 800 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory 802 or a program loaded from a storage device 808 into a random access memory 803. The random access memory 803 also stores various programs and data required for the operation of the electronic device 800. The processing unit 801, the read-only memory 802, and the random access memory 803 are interconnected via a bus 808. An input / output interface 805 is also connected to the bus 808.

[0157] Typically, the following devices can be connected to the input / output interface 805: input devices 808 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 808 including, for example, liquid crystal displays, speakers, vibrators, etc.; storage devices 808 including, for example, magnetic tapes, hard disks, etc.; and communication devices 809. Communication devices 809 allow electronic devices 800 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 8 An electronic device 800 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0158] Specifically, depending on the implementation of this technology, the process described in the above-referenced flowchart can be implemented as a computer software program. For example, one implementation includes a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. In such an implementation, the computer program can be downloaded and installed from a network via communication device 809, or installed from storage device 808, or installed from read-only memory 802. When the computer program is executed by processing device 801, it performs the functions defined in the method herein.

[0159] The names of the messages or information exchanged between multiple devices in this technical implementation are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0160] The electronic device and the computer program product integrated on the electronic device provided herein belong to the same inventive concept as the media content generation method provided in the above-described implementation. Technical details not described in detail herein can be found in the above-described implementation, and this document has the same beneficial effects as the above-described implementation.

[0161] This document provides a computer storage medium storing a computer program that, when executed by a processor, implements the media content generation method provided in the above-described embodiments. The computer storage medium provided herein and the media content generation method provided in the above-described embodiments belong to the same inventive concept. Technical details not described in detail herein can be found in the above-described embodiments, and this document has the same beneficial effects as the above-described embodiments.

[0162] It should be noted that the computer-readable medium described above can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. Computer-readable storage media can be, for example,—but not limited to—electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (ROM or flash memory), optical fibers, portable compact disk ROMs, optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0163] In this technology, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this technology, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, radio frequency, etc., or any suitable combination thereof.

[0164] Based on one or more implementations of this technology, Example 1 provides a media content generation method, comprising: obtaining first requested content; inputting the first requested content into a first model, the first model being distributed across multiple computing nodes; in response to a first execution operation, obtaining deployment change information of the first model, the deployment change information being determined based on first record information, the first record information being generated by the computing nodes; in response to a second execution operation, updating the first model based on the deployment change information, determining a first feature representation sequence through the computing nodes, the updating of the first model including updating the feature unit distribution and network structure distribution; and generating media content based on the first feature representation sequence.

[0165] Based on one or more implementations of this technology, Example 2 provides the method of Example 1, which optionally further includes: using the computing node, performing feature parsing on the first request content using the input layer of the first model to obtain a second feature representation sequence; using the feature processing layer of the first model to process the second feature representation sequence to obtain a third feature representation sequence; and in response to a third execution operation, using the third feature representation sequence as the second feature representation sequence, and returning to re-execute the step of using the feature processing layer of the first model to process the second feature representation sequence to obtain the third feature representation sequence.

[0166] Based on one or more implementations of this technology, Example 3 provides the method of Example 2. Optionally, the feature processing layer includes an attention module and an expert network module. The step of processing the second feature representation sequence to obtain the third feature representation sequence includes: using the attention module to perform attention processing on the second feature representation sequence to obtain an attention feature representation sequence; using the expert network module to assign the first feature unit to the first expert network for feature transformation according to the first allocation information, obtaining the feature transformation result output by the first expert network, and obtaining the third feature representation sequence according to the feature transformation result. The attention feature representation sequence contains the first feature unit, and the first allocation information records the allocation information of the first feature unit to the first expert network.

[0167] Based on one or more implementations of this technology, Example 4 provides the method of Example 1, which optionally involves obtaining, through the computing node, the processing time for the first model, the processing time being related to the processing of the first request content using the first model; in response to a first acquisition operation, obtaining second allocation information, the second allocation information being related to the allocation of feature units on the first model to a second expert network, the second expert network being included in the first model; and recording the processing time and the second allocation information to form the first recording information.

[0168] Based on one or more implementations of this technology, Example 5 provides the method of Example 1, wherein the step of determining the deployment change information based on the first record information includes: obtaining the second allocation information included in the first record information, and determining the feature deployment change information of the first model based on the second allocation information; determining the structural deployment change information of the first model based on the feature deployment change information and the second allocation information; and determining the deployment change information based on the feature deployment change information and the structural deployment change information.

[0169] Based on one or more implementations of this technology, Example Six provides the method of Example Five. Optionally, determining the feature deployment change information of the first model based on the second allocation information includes: obtaining the second allocation information from the first record information; determining the feature unit similarity matrix of the first model based on the second allocation information; determining the second feature unit associated with the computing node based on the feature unit similarity matrix, wherein the second feature unit is on the first model; and generating feature deployment change information based on the computing node and the second feature unit.

[0170] Based on one or more implementations of this technology, Example 7 provides the method of Example 5. Optionally, determining the structural deployment change information of the first model based on the feature deployment change information and the second allocation information includes: obtaining a first function, the first function being related to the running resources of the first model; determining the expert network demand degree of the computing node relative to the first model based on the feature deployment change information and the second allocation information; determining a third expert network associated with the computing node based on the expert network demand degree, the third expert network belonging to the first model, the third expert network satisfying a first condition, the first condition being that the function value of the first function is minimized; and generating structural deployment change information based on the computing node and the third expert network.

[0171] Based on one or more implementations of this technology, Example 8 provides the method of Example 1. Optionally, updating the first model according to the deployment change information includes: dividing a third feature unit to an associated computing node according to the feature deployment change information, wherein the third feature unit is on the first model; and deploying a fourth expert network to an associated computing node according to the structure deployment change information, wherein the fourth expert network belongs to the first model; wherein the feature deployment change information and the structure deployment change information are included in the deployment change information.

[0172] Based on one or more implementations of this technology, Example 9 provides the method of Example 1. Optionally, determining the first feature representation sequence through the computing node includes: performing attention processing on the fourth feature representation sequence through the computing node to obtain a fifth feature representation sequence, wherein the fourth feature representation sequence is the input content of the first model, the fifth feature representation sequence contains fourth feature units, and the fourth feature units are divided according to the deployment change information; allocating the fourth feature units to a fifth expert network for processing according to the second allocation information, obtaining the fifth feature units output by the fifth expert network, and obtaining the first feature representation sequence based on the fifth feature units, wherein the second allocation information is included in the first record information, and the fifth expert network is deployed according to the deployment change information.

[0173] Based on one or more implementations of this technology, Example 10 provides the method of Example 2. Optionally, the triggering condition for the first execution operation is that the number of iterations of feature processing of the first model reaches a first threshold; the triggering condition for the second execution operation is that the number of iterations is greater than the first threshold and the iteration has not ended; the triggering condition for the third execution operation is that the number of iterations is less than the first threshold.

[0174] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol, such as Hypertext Transfer Protocol, and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (LANs), wide area networks (WANs), the Internet (e.g., the Internet), and peer-to-peer networks, as well as any currently known or future-developed networks.

[0175] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0176] The aforementioned computer-readable medium carries one or more programs. When the electronic device executes the aforementioned one or more programs, the electronic device causes the electronic device to: acquire first requested content and input the first requested content into a first model, the first model being distributed across multiple computing nodes; in response to a first execution operation, acquire deployment change information of the first model, the deployment change information being determined based on first record information generated by the computing nodes; in response to a second execution operation, update the first model based on the deployment change information, determine a first feature representation sequence through the computing nodes, the updating of the first model including updating the feature unit distribution and network structure distribution; and generate media content based on the first feature representation sequence.

[0177] Computer program code for performing the operations of this technology can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages, as well as conventional procedural programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including local area networks (LANs) or wide area networks (WANs), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0178] The flowcharts and block diagrams in the accompanying figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various implementations of this technology. In this respect, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the figures. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0179] The modules or units described herein can be implemented in software or hardware. The name of a unit does not necessarily limit the unit itself; for example, a video generation request module can also be described as "a module for acquiring first text data".

[0180] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays, application-specific integrated circuits (ASICs), application-specific standard products (ASICs), systems-on-a-chip (SoCs), complex programmable logic devices, and so on.

[0181] In the context of this technology, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (flash memory), optical fibers, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0182] The above description is merely a preferred implementation of the technology and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this technology is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this technology that have similar functions.

[0183] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. Multitasking and parallel processing may be advantageous in certain contexts. Similarly, while some specific implementation details are included in the above discussion, these should not be interpreted as limitations on the scope of the technique. Certain features described in the context of a single implementation scenario can also be implemented in combination within that single implementation scenario. Conversely, various features described in the context of a single implementation scenario can also be implemented individually or in any suitable sub-combination in multiple implementation scenarios.

[0184] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for generating media content, comprising: Obtain the first request content and input the first request content into the first model, which is distributed across multiple computing nodes; In response to the first execution operation, deployment change information of the first model is obtained, the deployment change information is determined based on first record information, the first record information is generated by the computing node; In response to the second execution operation, the first model is updated according to the deployment change information, and the first feature representation sequence is determined through the computing node; Media content is generated based on the first feature representation sequence.

2. The method according to claim 1, further comprising: The computing node is used to perform feature parsing on the first request content using the input layer of the first model to obtain a second feature representation sequence. The feature processing layer of the first model is used to process the second feature representation sequence to obtain the third feature representation sequence; In response to the third execution operation, the third feature representation sequence is used as the second feature representation sequence, and the feature processing layer using the first model is re-executed to process the second feature representation sequence to obtain the third feature representation sequence.

3. The method according to claim 2, wherein the feature processing layer comprises an attention module and an expert network module. The step of processing the second feature representation sequence to obtain the third feature representation sequence includes: The attention module is used to perform attention processing on the second feature representation sequence to obtain an attention feature representation sequence; Using the expert network module, the first feature unit is assigned to the first expert network for feature transformation according to the first allocation information, the feature transformation result output by the first expert network is obtained, and the third feature representation sequence is obtained according to the feature transformation result. The attention feature representation sequence contains the first feature unit, and the first allocation information records the allocation information of the first feature unit to the first expert network.

4. The method according to claim 1, further comprising: The processing time for the first model is obtained through the computing node, and the processing time is related to the processing of the first request content using the first model; In response to the first acquisition operation, second allocation information is acquired, which is related to the allocation of feature units on the first model to the second expert network, and the second expert network is included in the first model; The processing time and the second allocation information are recorded to form the first recording information.

5. The method according to claim 1, wherein the step of determining the deployment change information based on the first record information includes: Obtain the second allocation information included in the first record information, and determine the feature deployment change information of the first model based on the second allocation information; Based on the feature deployment change information and the second allocation information, determine the structural deployment change information of the first model; The deployment change information is determined based on the feature deployment change information and the structure deployment change information.

6. The method according to claim 5, wherein determining the feature deployment change information of the first model based on the second allocation information includes: Obtain second allocation information from the first record information, and determine the feature unit similarity matrix of the first model based on the second allocation information; Based on the similarity matrix of the feature units, a second feature unit associated with the computing node is determined, and the second feature unit is on the first model; Based on the computing node and the second feature unit, feature deployment change information is generated.

7. The method according to claim 5, wherein determining the structural deployment change information of the first model based on the feature deployment change information and the second allocation information includes: Obtain the first function, which is related to the runtime resources of the first model; Based on the feature deployment change information and the second allocation information, the expert network demand degree of the computing node relative to the first model is determined; Based on the expert network demand degree, a third expert network associated with the computing node is determined. The third expert network belongs to the first model and satisfies a first condition, which is that the function value of the first function is minimized. Based on the computing nodes and the third expert network, structural deployment change information is generated.

8. The method according to claim 1, wherein updating the first model based on the deployment change information comprises: Based on the feature deployment change information, the third feature unit is assigned to the associated computing node, and the third feature unit is on the first model; According to the structural deployment change information, a fourth expert network is deployed to the associated computing nodes, and the fourth expert network belongs to the first model; The feature deployment change information and the structure deployment change information are included in the deployment change information.

9. The method according to claim 1, wherein determining the first feature representation sequence through the computing node comprises: Through the computing node, attention processing is performed on the fourth feature representation sequence to obtain the fifth feature representation sequence. The fourth feature representation sequence is the input content of the first model. The fifth feature representation sequence contains fourth feature units, which are divided according to the deployment change information. The fourth feature unit is assigned to the fifth expert network for processing according to the second allocation information, the fifth feature unit output by the fifth expert network is obtained, and the first feature representation sequence is obtained according to the fifth feature unit. The second allocation information is included in the first record information, and the fifth expert network is deployed according to the deployment change information.

10. The method according to claim 2, wherein the triggering condition for the first execution operation is that the number of iterations of feature processing performed by the first model reaches a first threshold; The triggering condition for the second execution operation is that the number of iterations is greater than the first threshold and the iteration has not ended; The trigger condition for the third execution operation is that the number of iterations is less than the first threshold.

11. A media content generation device, comprising: The acquisition module is used to acquire the first request content and input the first request content into the first model, which is distributed across multiple computing nodes. The first response module is used to respond to the first execution operation by obtaining deployment change information of the first model, wherein the deployment change information is determined based on first record information, and the first record information is generated by the computing node. The second response module is used to respond to the second execution operation, update the first model according to the deployment change information, determine the first feature representation sequence through the computing node, and update the first model includes updating the feature unit distribution and network structure distribution; The content generation module is used to generate media content based on the first feature representation sequence.

12. A computer program product comprising a computer program that, when executed by a processor, implements the media content generation method according to any one of claims 1-10.

Images