Hierarchical intelligent computing power supply system, model generation method, equipment and medium

Abstract

The embodiment of the invention provides a hierarchical intelligent calculation power supply system, a model generation method, equipment and a medium, and the system comprises a model algorithm design module which is used for generating a target logic model according to the model demands of a user; the training simulation and verification module is used for generating a target model and a virtual resource demand of the target model based on the target logic module; and the computing power base layout module is used for carrying out resource scheduling on the target model according to the virtual resource demand. Through the embodiment of the invention, the logic optimization and deployment optimization work of the large model can be decoupled, the cross-platform portability of the user logic model is ensured, the service iteration upgrading speed is accelerated, the computing power supply quality of the intelligent computing cloud platform is improved, and the intelligent computing service experience of the user in a large model scene is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of model generation, and in particular to a hierarchical intelligent computing power supply system, a model generation method, an electronic device and a computer-readable storage medium. Background Art

[0002] This year, with the continuous improvement of computing power, artificial intelligence (AI) technology has achieved a turning point and leapfrog development.

[0003] Currently, engineers can complete model deployment by designing models, writing model code, and configuring available resources. In actual applications, engineers need to spend a lot of time and energy to complete model design, model code writing, and available resource configuration. Summary of the invention

[0004] In view of the above problems, a hierarchical intelligent computing power supply system, a model generation method, an electronic device and a computer-readable storage medium are proposed to overcome the above problems or at least partially solve the above problems, including:

[0005] A hierarchical intelligent computing power supply system, the system comprising:

[0006] Model algorithm design module, used to generate target logic model according to user's model requirements;

[0007] A training, pushing, simulation and verification module, used to generate a target model and virtual resource requirements of the target model based on the target logic module;

[0008] The computing power base layout module is used to schedule resources for the target model according to the virtual resource requirements.

[0009] Optionally, the system further comprises:

[0010] The computing power system verification module is used to verify the target model.

[0011] Optionally, the computing power system verification module is also used to generate a scheduling optimization plan;

[0012] The computing power base layout module is also used to optimize the resource scheduling of the target model according to the scheduling optimization plan.

[0013] Optionally, the training, pushing, simulation and verification module is also used to generate a demand optimization plan;

[0014] The model algorithm design module is also used to display the demand optimization solution.

[0015] Optionally, the system further comprises a storage module;

[0016] The storage module is used to store the model parameters of the target model and the training data set for the target model.

[0017] Optionally, the storage module is used to store the own image uploaded by the user; and update the own image according to the model parameters.

[0018] Optionally, the model algorithm design module and the training, simulation and verification module are located on the model side; and the computing power base layout module is located on the computing power side.

[0019] The embodiment of the present invention further provides a method for generating a model, which is applied to the system as described above, and the method includes:

[0020] Obtain the target logical model generated according to the user's model requirements;

[0021] Based on the target logic module, a target model and a virtual resource requirement of the target model are generated; the virtual resource requirement is used by the system to perform resource scheduling on the target model.

[0022] An embodiment of the present invention further provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program implements the above-mentioned model generation method when executed by the processor.

[0023] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned model generation method is implemented.

[0024] The embodiments of the present invention have the following advantages:

[0025] In an embodiment of the present invention, the hierarchical intelligent computing power supply system includes: a model algorithm design module, which is used to generate a target logic model according to the user's model requirements; a training simulation and verification module, which is used to generate a target model based on the target logic module, as well as the virtual resource requirements of the target model; a computing power base layout module, which is used to schedule resources for the target model according to the virtual resource requirements. Through the embodiment of the present invention, the logic optimization and deployment optimization work of the large model can be decoupled, the cross-platform portability of the user's logical model can be guaranteed, the speed of service iteration and upgrade can be accelerated, the quality of the computing power supply of the intelligent computing cloud platform can be improved, and the user's intelligent computing service experience in the large model scenario can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solution of the present invention, the accompanying drawings required for use in the description of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0027] Figure 1 It is a structural schematic diagram of a hierarchical intelligent computing power supply system according to an embodiment of the present invention;

[0028] Figure 2 It is a structural schematic diagram of another hierarchical intelligent computing power supply system according to an embodiment of the present invention;

[0029] Figure 3 It is a structural diagram of another hierarchical intelligent computing power supply system according to an embodiment of the present invention;

[0030] Figure 4 is a schematic diagram of a data interaction process of an embodiment of the present invention;

[0031] Figure 5 It is a flowchart of the steps of a method for generating a model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] At present, artificial intelligence has entered the era of large models, and its use threshold has been greatly reduced, and the marginal cost is gradually approaching zero. However, the existing intelligent cloud platform scheduling optimization and cloud platform basic components have limitations. They make some restrictive assumptions on the applicability of simplified models, training settings, and networks.

[0034] For example, the communication scheduler makes assumptions about the model and training framework (simple layer-by-layer model and data-parallel training) or deployment method (parameter server-based), while the training framework for large models makes oversimplified assumptions about the nature of collective communication operations.

[0035] In addition, existing solutions are only point solutions in the optimization space and fail to collaboratively optimize resource scheduling and training and inference execution.

[0036] Therefore, there is a lot of room for optimization in the training and reasoning processes of large models on cloud platforms, and there is a lack of a clear architecture to promote the evolution of optimization solutions at different levels.

[0037] The present invention is aimed at the scenarios of large-model training and reasoning of the intelligent computing cloud platform. Through the demand analysis of the large-model design, deployment, and operation links, a hierarchical cloud platform intelligent computing power supply system is creatively proposed. The invention can decouple the logic optimization and deployment optimization of large models, ensure the cross-platform portability of user logical models, accelerate the speed of service iteration and upgrade, improve the quality of intelligent computing cloud platform computing power supply, and improve the user's intelligent computing service experience in large model scenarios. The present invention is not only applicable to (GPUGraphics Processing Unit, graphics processor) graphics processor clusters, but also to other intelligent computing computing power supply clusters constructed with NPU (Neural Processing Unit) neural processing unit chips specially designed for artificial intelligence and machine learning.

[0038] For details, please refer to Figure 1 , Figure 1 A structural schematic diagram of a hierarchical intelligent computing power supply system according to an embodiment of the present invention is shown.

[0039] like Figure 1 As shown, the hierarchical intelligent computing power supply system 10 may include the following modules:

[0040] The model algorithm design module 110 is used to generate a target logic model according to the user's model requirements;

[0041] The training, pushing, simulation and verification module 120 is used to generate a target model and virtual resource requirements of the target model based on the target logic module;

[0042] The computing power base layout module 130 is used to schedule resources for the target model according to the virtual resource requirements.

[0043] In practical applications, the hierarchical intelligent computing power supply system 10 may be composed of a model algorithm design module 110, a training simulation and verification module 120, and a computing power base layout module 130. Among them:

[0044] The model algorithm design module 110 can be used to obtain the user's model requirements and generate a target logic model based on the user's model requirements; the target logic model can be expressed in the form of a high-level programming language such as Python.

[0045] Exemplarily, the model algorithm design module 110 may also support the user to complete the model design by dragging a graphic, and convert a large model in a graphical form into a general model description language such as Python as a target logical model.

[0046] After generating the target logic model, the model algorithm design module 110 may send the target logic model to the training simulation and verification module 120 for the next step.

[0047] Specifically, after receiving the target logic model, the training, simulation and verification module 120 can compile and analyze the target logic model and generate a target model; in addition, it can also generate virtual resource requirements for the target model.

[0048] Among them, virtual resource requirements can refer to the virtual resources required for target model deployment.

[0049] For example: GPU resources, CPU resources, memory, etc., which are not limited in the embodiment of the present invention.

[0050] In actual applications, after determining the target model and virtual resource requirements, the computing power base layout module 130 can open computing power resources for the target model according to the virtual resource requirements, and complete resource scheduling for the target model, so that when the target model is deployed, sufficient resources can be called to complete the deployment of the model.

[0051] In an embodiment of the present invention, the hierarchical intelligent computing power supply system 10 includes: a model algorithm design module 110, which is used to generate a target logic model according to the user's model requirements; a training simulation and verification module 120, which is used to generate a target model and virtual resource requirements of the target model based on the target logic module; a computing power base layout module 130, which is used to schedule resources for the target model according to the virtual resource requirements.

[0052] Through the embodiments of the present invention, the logic optimization and deployment optimization work of large models can be decoupled, the cross-platform portability of user logic models can be ensured, the speed of service iteration and upgrading can be accelerated, the quality of computing power supply of the intelligent computing cloud platform can be improved, and the user's intelligent computing service experience in large model scenarios can be improved.

[0053] In one embodiment of the present invention, Figure 2 As shown, the above-mentioned hierarchical intelligent computing power supply system 10 can also exist in another form.

[0054] Specifically, in addition to the model algorithm design module 110, the training simulation and verification module 120, and the computing power base layout module 130, the hierarchical intelligent computing power supply system 10 may also include the following modules:

[0055] The computing system verification module 140 is used to verify the target model.

[0056] In some feasible embodiments, the complexity of large models and their training is relatively high. Even after the above-mentioned preliminary scheduling and deployment is completed, the state of the system during actual operation is difficult to guarantee, and may even deviate greatly from the prediction of the theoretical model.

[0057] In order to provide SLA (Service Level Agreement) guarantees to customers, after the initial deployment, the computing power system verification module 140 can start distributed programs such as cuda-test and nccl-test for computing and communication testing in the background to test the target model and obtain the actual operating performance parameters of the target model; further, the computing power system verification module 140 can verify the target model based on the operating performance parameters to determine whether the generated target model and the resources scheduled for the target model are appropriate.

[0058] In one embodiment of the present invention, the computing power system verification module 140 is also used to generate a scheduling optimization plan;

[0059] The computing power base layout module 130 is also used to optimize the resource scheduling of the target model according to the scheduling optimization plan.

[0060] In some feasible embodiments, the computing power system verification module 140 can also generate a scheduling optimization plan based on the actual operating performance parameters of the target model to optimize the resource scheduling of the computing power base layout model.

[0061] Specifically, when verifying that there is unreasonable resource scheduling for the target model, the computing power system verification module 140 can generate a scheduling optimization method for the target model and feed it back to the computing power base layout module 130.

[0062] The computing power base layout module 130 can optimize the resource scheduling of the target model based on the scheduling optimization plan; specifically, the computing power base layout module 130 can redetermine the resource allocation for the target model to complete the optimization of resource layout and scheduling.

[0063] In one embodiment of the present invention, the training, pushing, simulation and verification module 120 is also used to generate a demand optimization solution;

[0064] The model algorithm design module 110 is also used to display demand optimization solutions.

[0065] In some feasible embodiments, the training, simulation and verification module 120 may also optimize the training and reasoning framework for the target model during the process of generating the target model, and feedback the required optimization plan to the model algorithm design module 110 so that the user can reconstruct the target model on the model algorithm design module 110.

[0066] Exemplarily, the training, simulation and verification module 120 can provide feedback to the model algorithm design module 110, helping users design better models and refactor Python codes.

[0067] The output of the training, simulation and verification module 120 is a computational communication diagram of the large model training and a list of virtual resources, that is, a large model that is preliminarily optimized and clarifies the virtual resource requirements.

[0068] In one embodiment of the present invention, the system further includes a storage module;

[0069] The storage module is used to store model parameters of the target model and a training data set for the target model.

[0070] In some feasible embodiments, the hierarchical intelligent computing power supply system 10 may include a storage module in addition to the model algorithm design module 110, the training simulation and verification module 120 and the computing power base layout module 130.

[0071] Exemplarily, the storage module can be used for various data in the entire process of building the target model. Specifically:

[0072] The storage module may be used to store the model parameters of the target model generated by the training, simulation and verification module 120 .

[0073] In addition, the storage module can also store the training data set for the target model uploaded by the user.

[0074] The training data set may include a training data set and a test data set, which is not limited in the embodiment of the present invention.

[0075] In one embodiment of the present invention, the storage module is used to store the own image uploaded by the user; and to update the own image according to the model parameters.

[0076] In some feasible embodiments, the storage module can store custom images uploaded by users; exemplary, custom images refer to virtual machine images created and managed by users according to their own needs. These images usually contain specific software configurations, libraries, data sets, and environment settings, and users can reuse these images to start virtual machine instances, thereby saving configuration time and ensuring the consistency of the environment.

[0077] In addition, after determining the model parameters of the target model, the training, simulation and verification module 120 can upload the model parameters of the target model to the storage module to update its own image, which is not limited in this embodiment of the present invention.

[0078] In one embodiment of the present invention, the model algorithm design module 110 and the training simulation and verification module 120 are located on the model side; and the computing power base layout module 130 is located on the computing power side.

[0079] In some feasible embodiments, the model algorithm design module 110 and the training, simulation and verification module 120 can be deployed on the model side; the computing power base layout module 130 can be deployed on the computing power side; through decoupling, and connecting the various sides through TCP / IP or a high-performance network; the user can concentrate on the logical design of the model; and through decoupling, the model design and resource scheduling can be carried out in different environments, so that the user can complete the preparation of the image file in his own environment or a third-party cloud environment, and make the generated target model portable.

[0080] For example, because the number of model parameters of large models is in the billions, the amount of training data and the storage of intermediate variables generated by training inference are difficult to meet with the high bandwidth memory (HBM) of a single GPU card or the CPU memory of a single node. Therefore, distributed training has become the basic technology for the evolution of large models.

[0081] The communication of large models is mainly to update the gradients obtained by distributed computing to the models on each node to achieve parameter synchronization for one iteration. Common communication terms include: Broadcast, Scatter, Reduce, AllReduce, Gather, AllGather, ReduceScatter.

[0082] Based on the above-mentioned collective communication primitives, the existing intelligent cloud platform scheduling optimization and execution planner will convert the logic design of the large model to be deployed (written by developers based on the front-end model programming framework) into a computing-communication directed acyclic graph through the model compiler during system initialization, and allocate underlying computing resources to support the training and reasoning process of the large model.

[0083] The hierarchical intelligent computing power supply system 10 of the embodiment of the present invention is based on the design and development life cycle of the big model, utilizes the network topology information, node resource quota information and host information independently maintained by the cloud platform, reconstructs the architecture design of the computing power supply system of the big model, and creatively proposes a hierarchical cloud platform intelligent computing power supply method and system. The specific implementation steps are briefly described as follows:

[0084] like Figure 3 As shown, the first layer is the model algorithm design layer, namely the model algorithm design module 110. This layer mainly includes a model editing framework composed of programming framework interfaces for various large model developments and a visual model design interface.

[0085] These framework interfaces encapsulate the sub-functional modules (such as various machine learning operators) that constitute the large model.

[0086] Large model developers can perform top-level logic design based on these modules and import data sets in various formats, dividing them into training data sets and test sets as needed.

[0087] At this layer, users focus on model construction without having to consider specific runtime strategies such as data parallelism, pipeline parallelism, or tensor parallelism.

[0088] The output of this layer is the logical model written by the user. This product is generally expressed in the form of high-level programming languages ​​​​such as Python. It can also support users to complete model design by dragging and dropping graphics, and convert large graphical models into general model description languages ​​​​such as Python.

[0089] The second layer is the training, inference, simulation and verification layer, namely the training, inference, simulation and verification module 120. This layer mainly includes a variety of training compilers, internal implementation of the large model training framework, and a high-order logical cost model for training reasoning.

[0090] The training compiler can decompose the model training process into a directed acyclic computational communication graph.

[0091] The training framework performs high-level logical modeling on the training process of large language models.

[0092] By using a high-order logical cost model, the training framework will formally estimate the computational and storage requirements of large model training, and introduce data parallelism, pipeline parallelism, or (and) tensor parallelism to optimize the training process of large models.

[0093] This layer focuses on using software tools to perform formal analysis of the model and its training process based on high-order logic, and to estimate the virtual resource requirements of the training process (the number of channels with a specific virtual communication topology, the average channel bandwidth, the average channel latency, the number of computing cores, and the size of the storage space, etc.).

[0094] This layer also provides feedback to the first layer, helping users design better models and refactor Python code.

[0095] The output of this layer is the computational communication graph of the large model training and the virtual resource list (the virtual resource list is the virtual resource demand), that is, the large model that preliminarily optimizes and clarifies the virtual resource demand.

[0096] The third layer is the computing power base layout layer, that is, the computing power base layout module 130. This layer mainly includes resource schedulers and layout optimizers such as the PaaS (Platform as a Service) scheduler, virtual resource layouter and IaaS (Infrastructure as a Service) scheduler of the cloud platform.

[0097] The previous layer defines the specific virtual resource requirements and implements optimization at the high-level logic level, but when virtual resources are laid out to the underlying physical resources, problems such as contention and over-subscription may arise.

[0098] The PaaS layer container scheduler first allocates a number of working nodes in multiple Pods based on the user's demand for virtual computing resources and the coarse-grained affinity principle.

[0099] On this basis, the virtual resource planner performs graph modeling of the underlying hardware computing resources, communication link topology connection relationships and storage resources.

[0100] By mapping virtual resources to underlying hardware resources through the layout device, hardware resource conflicts during the training process are reduced.

[0101] Finally, the IaaS scheduler is responsible for scheduling the layout mapping to specific hardware resources. The output of this layer is the resource deployment model and the large model of the completed deployment.

[0102] The fourth layer is the computing power system verification layer, namely the computing power system verification module 140. This layer mainly includes test programs for computing power and communication capabilities and post-simulation programs to assist prediction.

[0103] Due to the high complexity of large models and their training, even after the above preliminary scheduling and deployment is completed, the status of the system during actual operation is difficult to guarantee, and may even deviate greatly from the predictions of the theoretical model.

[0104] In order to provide customers with SLA (Service Level Agreement) guarantees, after the initial deployment, the cloud platform will start distributed programs such as cuda-test and nccl-test for computing and communication testing in the background to obtain actual operating performance parameter estimates.

[0105] Based on this set of measured parameters, we then update our estimate of the model training capabilities based on the layouter's modeling of the system.

[0106] This layer provides feedback to the third layer, performs deployment optimization for models whose actual performance is far from expected, and allows the layouter to rearrange and the controller to reschedule. The output of this layer is a deployment optimization plan and a clear SLA.

[0107] The above describes the hierarchical intelligent computing power supply system 10 from the perspective of architecture. Next, we will Figure 4 Take the example of the system, and then explain the behavior of the system from the user's perspective (workflow):

[0108] First, based on the control panel and high-performance network, the intelligent computing user first opens and logs in to the account of the hierarchical intelligent computing power supply system 10 (referred to as the intelligent computing cloud platform) on the user side, and calls the model algorithm design interface to complete the large model design, upload the data set and the own image to the intelligent computing storage on the storage side for cloud storage, and use the model algorithm design interface to build a customized large model.

[0109] Afterwards, the logical model needs to be optimized in combination with the user-specified training framework (such as DeepSpeed ​​or Megatron). This part can introduce the reconstruction of the existing model code to optimize the training and reasoning performance of the original logical model, and finally obtain an image deployment file containing model code, training scripts, etc. (corresponding to the second layer of the above architecture).

[0110] The image file is the logical design of the model, and preliminary performance optimization has been performed on the logical cost model. The product is portable.

[0111] Therefore, users can also prepare the above image files in their own environment or a third-party cloud environment.

[0112] After compilation and analysis, the second layer can optimize the training and reasoning framework of the model and feed back the optimization to the first layer so that the first layer can reconstruct the logical model.

[0113] After the first layer completes the reconstruction, it can notify the second layer that the reconstruction is complete; when the second layer is optimized, the model parameters of the optimized model can be written to the cloud storage of the intelligent computing storage to update the model image.

[0114] After that, the user submits the above logical model and training data set to the Zhisuan Cloud Platform, and the virtual resource planner and scheduler of the Zhisuan Cloud Platform will complete the allocation of the user's virtual resource needs.

[0115] In the third layer, self-developed underlying intelligent computing components such as communication libraries are provided, and can cover the underlying components specified in the container image (such as Megatron and Pytroch are NCCL by default), so that customers are unaware of the underlying communications, cuda versions, etc., reducing the adaptation workload.

[0116] Specifically, the third layer can complete the virtual resource layout and scheduling, and open up computing resources for the target model.

[0117] Then, the fourth layer can perform computing power testing and verification on the target model and feedback the test results to the third layer.

[0118] Specifically, the intelligent computing platform will verify the performance of network, computing and storage resources through test programs, and predict the training time cost of the model through post-simulation services.

[0119] The third layer can perform layout and scheduling optimization, redetermine the computing resources deployed for the target model, and notify the cloud storage to update.

[0120] At this point, the third layer can pull the updated model image and data set, notify the user that the model deployment is complete, and provide a reference SLA.

[0121] When intelligent computing users start large-model training and reasoning, they can call the intelligent computing cloud platform interface; the third layer can call the underlying components, monitor and ensure the operating stability of the target model.

[0122] After the target model is trained, the model side can notify the training completion, and the computing power side can back up the model parameters to the intelligent computing storage on the storage side, which can store the trained model parameters. On the user side, intelligent computing users can carry out intelligent computing services based on the target model.

[0123] Based on the above-mentioned hierarchical intelligent computing power supply system, the embodiment of the present invention also provides a method for generating a model. For details, please refer to Figure 5 .

[0124] Figure 5 A flow chart showing the steps of a method for generating a model according to an embodiment of the present invention is shown; Figure 5 As shown, the method for generating the model may include the following steps:

[0125] Step 501: Obtain a target logical model generated according to the user's model requirements.

[0126] The model algorithm design module can be used to obtain the user's model requirements and generate a target logic model based on the user's model requirements; the target logic model can be expressed in the form of a high-level programming language such as Python.

[0127] Exemplarily, the model algorithm design module may also support users to complete model design by dragging graphics, and convert large models in graphical form into general model description languages ​​such as Python as the target logical model.

[0128] After generating the target logic model, the model algorithm design module can send the target logic model to the training simulation and verification module for the next step.

[0129] Step 502: Generate a target model and virtual resource requirements of the target model based on the target logic module; the virtual resource requirements are used by the system to schedule resources for the target model.

[0130] After receiving the target logic model, the simulation and verification module can compile and analyze the target logic model and generate the target model; in addition, it can also generate virtual resource requirements for the target model.

[0131] In actual applications, after determining the target model and virtual resource requirements, the computing power base layout module can open computing power resources for the target model according to the virtual resource requirements and complete resource scheduling for the target model, so that when the target model is deployed, sufficient resources can be called to complete the model deployment.

[0132] In the embodiment of the present invention, a target logic model generated according to the user's model requirements is obtained; based on the target logic module, a target model and a virtual resource requirement of the target model are generated; the virtual resource requirement is used by the system to schedule resources for the target model. Through the embodiment of the present invention, the logic optimization and deployment optimization work of the large model can be decoupled, the cross-platform portability of the user logic model can be guaranteed, the speed of service iteration and upgrade can be accelerated, the quality of the computing power supply of the intelligent computing cloud platform can be improved, and the user's intelligent computing service experience in the large model scenario can be improved.

[0133] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0134] An embodiment of the present invention further provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program implements the above-mentioned model generation method when executed by the processor.

[0135] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned model generation method is implemented.

[0136] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0137] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0138] Those skilled in the art will appreciate that the embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the embodiments of the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0139] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0140] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0141] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0142] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0143] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0144] The above provides a detailed introduction to a hierarchical intelligent computing power supply system, a model generation method, an electronic device and a computer-readable storage medium. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A hierarchical intelligent computing power supply system, characterized in that: The system comprises: Model algorithm design module, used to generate target logic model according to user's model requirements; A training, pushing, simulation and verification module, used to generate a target model and virtual resource requirements of the target model based on the target logic module; The computing power base layout module is used to schedule resources for the target model according to the virtual resource requirements.

2. The system according to claim 1, characterized in that The system further comprises: The computing power system verification module is used to verify the target model.

3. The system according to claim 2, characterized in that The computing power system verification module is also used to generate a scheduling optimization plan; The computing power base layout module is also used to optimize the resource scheduling of the target model according to the scheduling optimization plan.

4. The system according to claim 1, characterized in that The training, pushing, simulation and verification module is also used to generate a demand optimization plan; The model algorithm design module is also used to display the demand optimization solution.

5. The system according to claim 1, characterized in that The system also includes a storage module; The storage module is used to store the model parameters of the target model and the training data set for the target model.

6. The system according to claim 5, characterized in that The storage module is used to store the own image uploaded by the user; and update the own image according to the model parameters.

7. The system according to claim 1, characterized in that The model algorithm design module and the training simulation and verification module are located on the model side; the computing power base layout module is located on the computing power side.

8. A method for generating a model, characterized in that: Applied to the system according to any one of claims 1 to 7, the method comprises: Obtain the target logical model generated according to the user's model requirements; Based on the target logic module, a target model and a virtual resource requirement of the target model are generated; the virtual resource requirement is used by the system to perform resource scheduling on the target model.

9. An electronic device, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and capable of running on the processor, wherein the computer program implements the method for generating a model as claimed in claim 8 when executed by the processor.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for generating a model as claimed in claim 8 is implemented.

Images