Large model network dynamic expansion method and device of intelligent computing center cloud platform providing computing power resources

Abstract

The application provides a large model network dynamic expansion method and device of an intelligent computing center cloud platform providing computing power resources, comprising: receiving a sample question set input by a user; training a large model based on the sample question set to obtain a trained large model; the large model is provided with a jump gating module and a cycle gating module; the jump gating module and the cycle gating module are used for learning based on the sample question set in the training process of the large model to gradually learn an optimal dynamic decision rule; the trained large model is used for receiving a to-be-processed question input by a user, reasoning the to-be-processed question based on the jump gating module, the cycle gating module and the optimal dynamic decision rule, dynamically expanding multiple network layers of the large model in the reasoning process, and obtaining an output result based on the expanded multiple network layers. The number of network layers is dynamically expanded in the training and reasoning process of the large model, which can improve the reasoning performance of the large model while reducing the consumption of computing power resources.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent computing centers, smart computing centers, and computing infrastructure, specifically to a method and apparatus for dynamically expanding a large-scale network of an intelligent computing center cloud platform that provides computing resources. Background Technology

[0002] With the rapid development of artificial intelligence technology, "intelligent computing centers" and "smart computing centers" have emerged.

[0003] An "intelligent computing center" refers to a facility that provides the necessary computing power, data, and algorithms for artificial intelligence applications (such as deep learning model development, model fine-tuning, and model inference) by utilizing large-scale heterogeneous computing resources, including general-purpose and intelligent computing power. Intelligent computing centers encompass facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enablement.

[0004] "Intelligent computing center" includes, but is not limited to, "intelligent computing center".

[0005] "Intelligent computing center" or artificial intelligence computing center is a type of computing infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications, based on artificial intelligence theory and adopting artificial intelligence computing architecture.

[0006] "Computing power" is the core of "intelligent computing center" and "smart computing center". It is the ability of computer equipment or computing / data center to process information. It is the ability of computer hardware and software to work together to perform a certain computing requirement. It is the computing power to achieve the target output by processing information data. It is a new type of productivity that integrates information computing power, network carrying capacity and data storage capacity. It mainly provides services to society through computing power infrastructure.

[0007] In the current field of artificial intelligence, especially in tasks such as natural language processing and computer vision, large models have become the mainstream in research and application. However, these large models often have a huge number of parameters, resulting in the consumption of a large amount of computing resources during deployment and use. Furthermore, the existing Transformer network architecture has a linear and static structure, meaning that the number of network layers in large models designed based on the Transformer architecture is fixed during the training and use phases, leading to significant computing resource consumption during inference and training as well. This not only wastes computing resources but also increases the operating cost of large models, limiting their widespread application.

[0008] In summary, since the emergence of intelligent computing centers, large models, due to their massive number of parameters and fixed network layer design, have led to a waste of computing resources and an increase in operating costs during large model inference and training. This problem has become an urgent technical issue to be addressed. Summary of the Invention

[0009] This invention provides a method and apparatus for dynamically expanding large model networks in an intelligent computing center cloud platform that provides computing power resources. This addresses the technical problem that, since the emergence of intelligent computing centers, large models have resulted in wasted computing power resources and increased operating costs during large model inference and training due to their large number of parameters and fixed network layer design.

[0010] To solve the above-mentioned technical problems, the present invention is implemented as follows:

[0011] In a first aspect, the present invention provides a method for dynamically expanding a large-scale model network of an intelligent computing center cloud platform that provides computing power resources, the method comprising:

[0012] Step S1: Receive the sample question set input by the user;

[0013] Step S2: Based on the sample problem set, train the large model to obtain the trained large model; wherein, the large model is equipped with a jump gate module and a loop gate module;

[0014] The jump gate module and the loop gate module are used to learn based on the sample problem set during the training process of the large model, so as to gradually learn the optimal dynamic decision rule.

[0015] The trained large model is used to receive the user input of the problem to be processed, and to reason about the problem based on the jump gating module, the loop gating module and the optimal dynamic decision rule. During the reasoning process, it dynamically expands its multiple network layers and obtains the output result based on the expanded multiple network layers.

[0016] The optimal dynamic decision rule is used to determine, based on the problem to be solved, the network layers that need to be skipped and the network layers that need to be executed cyclically among the multiple network layers.

[0017] Optionally, step S2 includes:

[0018] Step S21: Based on the sample problem set, train the large model using a preset loss function, with the goal of minimizing the loss function, to obtain the trained large model;

[0019] The loss function is used to measure the difference between the output of the large model and the standard answer, and / or to measure the use of computing resources during the training of the large model.

[0020] Optionally, the architecture of the trained large model is a Transformer architecture, and the trained large model of the Transformer architecture includes at least an embedding layer and the multiple network layers;

[0021] The jump gating module is used to receive the output information of the embedding layer, and output a jump index based on the output information of the embedding layer and the optimal dynamic decision rule. The jump index is used to indicate the jump execution state of each of the plurality of network layers, wherein the jump execution state includes skipping the current network layer and activating the current network layer.

[0022] The loop gating module is used to receive the output information of the embedding layer, and output a loop index based on the output information of the embedding layer and the optimal dynamic decision rule, wherein the loop index is used to indicate the loop information of each network layer in the plurality of network layers;

[0023] The cycle information includes any one of the following: the current network layer belongs to a cycle group and the number of cycles in the cycle group; the current network layer does not belong to a cycle group.

[0024] Wherein, if the current network layer is selected independently or not selected, the loop information corresponding to the current network layer is: the current network layer does not belong to the loop group; if the current network layer is selected, and at least one of the network layers adjacent to the current network layer is also selected, the loop information corresponding to the current network layer is: the current network layer belongs to the loop group and the number of loops of the loop group;

[0025] The plurality of network layers are used to determine whether to perform a jump based on the jump index and whether to perform a loop based on the loop index.

[0026] Optionally, the number of iterations is related to the complexity of the problem to be processed; the higher the complexity of the problem to be processed, the greater the number of iterations.

[0027] Optionally, if there is a skip network layer in the loop group that skips the current network layer, then the difference between the total number of network layers in the loop group and the number of skip network layers is determined. If the difference is greater than or equal to 2, then the remaining network layers in the loop group, excluding the skip network layers, are looped.

[0028] Secondly, the present invention provides a dynamic expansion device for a large-scale model network of an intelligent computing center cloud platform that provides computing power resources, the device comprising:

[0029] The receiving module is used to perform step S1: receiving a set of sample questions input by the user;

[0030] The execution module is used to execute step S2: based on the sample problem set, train the large model to obtain the trained large model; wherein, the large model is equipped with a jump gate module and a loop gate module;

[0031] The jump gate module and the loop gate module are used to learn based on the sample problem set during the training process of the large model, so as to gradually learn the optimal dynamic decision rule.

[0032] The trained large model is used to receive the user input of the problem to be processed, and to reason about the problem based on the jump gating module, the loop gating module and the optimal dynamic decision rule. During the reasoning process, it dynamically expands its multiple network layers and obtains the output result based on the expanded multiple network layers.

[0033] The optimal dynamic decision rule is used to determine, based on the problem to be solved, the network layers that need to be skipped and the network layers that need to be executed cyclically among the multiple network layers.

[0034] Optionally, step S2 includes:

[0035] Step S21: Based on the sample problem set, train the large model using a preset loss function, with the goal of minimizing the loss function, to obtain the trained large model;

[0036] The loss function is used to measure the difference between the output of the large model and the standard answer, and / or to measure the use of computing resources during the training of the large model.

[0037] Thirdly, the present invention provides a server comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of a method for dynamically expanding a large-scale network of an intelligent computing center cloud platform that provides computing resources as described in the first aspect above.

[0038] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for dynamically expanding a large-scale network of an intelligent computing center cloud platform that provides computing resources, as described in the first aspect above.

[0039] Fifthly, the present invention provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of a method for dynamically expanding a large model network of an intelligent computing center cloud platform that provides computing resources as described in the first aspect above.

[0040] In this invention, the large model is equipped with a skip gating module and a loop gating module. During training, the large model learns from a sample problem set to gradually learn the optimal dynamic decision rule. The skip gating module allows the large model to intelligently skip unnecessary network layers during inference, based on the problem to be addressed and the optimal dynamic decision rule, thereby reducing computational load and resource consumption. The loop gating module, on the other hand, can repeatedly execute certain key network layers when needed (e.g., when the problem to be addressed is complex), enhancing the large model's understanding and processing capabilities for complex problems. This ability to dynamically expand the number of network layers in the large model enables the trained model to adaptively adjust the network layers according to the problem to be addressed (the actual input), thus saving computational resources, avoiding waste, and improving the utilization rate of computational resources while ensuring the inference performance of the large model. Attached Figure Description

[0041] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0042] Figure 1 A flowchart illustrating a method for dynamically expanding a large-scale network of an intelligent computing center cloud platform that provides computing power resources, as provided by this invention.

[0043] Figure 2 This invention provides a network hierarchy diagram of a large model of the existing Transformer architecture;

[0044] Figure 3 This invention provides a network hierarchy structure diagram of a large model of an intelligent computing center cloud platform that provides computing power resources;

[0045] Figure 4 The structural block diagram of a large-scale network dynamic expansion device for an intelligent computing center cloud platform that provides computing power resources is provided by the present invention.

[0046] Figure 5 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The technical terms involved in this invention will be briefly explained below.

[0049] The “computing power” mentioned in this invention refers to: the ability of computer equipment or computing / data center to process information; the ability of computer hardware and software to work together to perform a certain computing requirement; the computing power to achieve the target result output by processing information data; and a new type of productivity that integrates information computing power, network carrying capacity, and data storage capacity, mainly providing services to society through computing power infrastructure.

[0050] The "computational power" (CP) described in this invention refers to the ability of a data center server to process data and output results. It is a comprehensive indicator of a data center's computing power, encompassing general computing power, supercomputing power, and intelligent computing power. The commonly used unit of measurement is floating-point operations per second (FLOPS, 1 EFLOPS = 10^18 FLOPS), with higher values ​​indicating stronger overall computing power. It is estimated that 1 EFLOPS is approximately the computing power output of 5 Tianhe-2A supercomputers, 500,000 mainstream server CPUs, or 2 million mainstream laptops. The calculation formula is: CP = General CP + Intelligent CP + Super CP

[0051] The "Network Power" (NP) mentioned in this invention refers to the performance of data transmission capabilities of computing facilities, which includes comprehensive capabilities such as network architecture, network bandwidth, transmission latency, intelligent management and scheduling, and involves network transmission within and between data centers. It is a comprehensive indicator for measuring network transmission scheduling capabilities.

[0052] The "Storage Power" (SP) described in this invention refers to the comprehensive capabilities of a data center in four aspects: data storage capacity, performance, security and reliability, and green and low-carbon operation. It is a comprehensive indicator for measuring the data storage capacity of a data center, including external storage devices such as storage arrays and internal storage devices in servers. The commonly used unit of measurement for storage capacity is exabytes (EB, 1EB = 2^60 bytes), the commonly used unit of measurement for performance is the number of read / write operations per second per unit capacity (IOPS / TB), and the disaster recovery ratio is an important indicator of security and reliability.

[0053] The "computing infrastructure" mentioned in this invention refers to a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity, enabling centralized computing, storage, transmission, and application of information.

[0054] The "new information infrastructure" mentioned in this invention refers to network infrastructure such as 5G networks, fiber optic broadband networks, backbone networks, international communication networks, and satellite internet; computing infrastructure such as data centers, general computing centers, intelligent computing centers, and supercomputing centers; and new technology facilities such as artificial intelligence, blockchain, and quantum computing.

[0055] The “computing power” mentioned in this invention includes: “general computing power”, “intelligent computing power” and “supercomputing power”.

[0056] The "general computing power" mentioned in this invention refers to the computing power provided by servers based on CPU (Central Processing Unit) chips, which is used to support basic general computing such as cloud computing and edge computing.

[0057] The "intelligent computing power" mentioned in this invention refers to: a computing platform deployed on a large scale based on dedicated chips such as GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), and ASIC (Application Specific Integrated Circuit) for various artificial intelligence innovative applications, such as natural language processing and machine vision.

[0058] The “supercomputing power” mentioned in this invention refers to the computing power provided by high-performance computing clusters such as supercomputers. It utilizes the centralized computing resources of multiple computer systems working in parallel and uses a dedicated operating system to handle extremely complex or data-intensive problems. It is mainly used for computing in cutting-edge scientific fields, such as planetary simulation, drug molecule design, and gene analysis.

[0059] The "intelligent computing center" described in this invention refers to a facility that, through the use of large-scale heterogeneous computing resources, including general-purpose computing power (CPU) and intelligent computing power (GPU, FPGA, ASIC, etc.), primarily provides the necessary computing power, data, and algorithms for artificial intelligence applications (such as deep learning model development, model fine-tuning, and model inference). The intelligent computing center encompasses facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enablement.

[0060] The "intelligent computing center cloud platform" described in this invention refers to a cloud computing platform that integrates hardware and software resources of an intelligent computing center.

[0061] The "intelligent computing center" mentioned in this invention includes, but is not limited to, "smart computing center".

[0062] The "intelligent computing center" mentioned in this invention, also known as an artificial intelligence computing center, is a type of computing infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications, based on artificial intelligence theory and adopting an artificial intelligence computing architecture.

[0063] The "computing center" mentioned in this invention refers to a facility that is mainly composed of infrastructure such as wind, thermal, hydro, and electricity, and IT hardware and software equipment, and has computing power, carrying capacity, and storage capacity, including general data centers, intelligent computing centers, supercomputing centers, etc.

[0064] The "supercomputing center" mentioned in this invention refers to a supercomputing data center, which is a data center based on supercomputers or large-scale computing clusters. It can provide large-scale computing, storage and network services and is widely used in aerospace, defense, oil exploration, climate modeling and genome sequencing and other application scenarios.

[0065] The “computing resources” mentioned in this invention refer to the technologies and facilities required for the development of the digital society that have the ability to compute, transmit, store and apply information, including but not limited to computing resources such as CPUs and GPUs, network resources such as switches and routers, storage resources such as storage arrays and distributed storage, security resources such as firewalls and intrusion detection systems, and supporting and guaranteeing resources such as wind, fire, water and electricity.

[0066] The "large model" mentioned in this invention refers to a large-scale language model (LLM), which is a language model with a large number of parameters. It is designed to understand and generate human language, and can be fine-tuned through a large amount of text data to perform a wide range of tasks, including text summarization, translation, and sentiment analysis.

[0067] The "Poisson distribution" described in this invention is a discrete probability distribution commonly found in statistics and probability, suitable for describing the probability distribution of the number of times a random event occurs per unit time.

[0068] The "softmax function" described in this invention is a mathematical function that normalizes a vector of arbitrary real values ​​(usually unnormalized fractions or log odds) into a probability distribution.

[0069] Figure 1 This invention illustrates a method for dynamically expanding a large-scale model network of an intelligent computing center cloud platform that provides computing power resources, as provided by the present invention. Figure 1 As shown, the method includes:

[0070] Step S1: Receive the sample question set input by the user;

[0071] Step S2: Based on the sample problem set, train the large model to obtain the trained large model; the large model is equipped with a jump gating module and a loop gating module.

[0072] It should be noted that the jump gating module and the loop gating module are used to learn based on the sample problem set during the training of the large model, so as to gradually learn the optimal dynamic decision rule. The trained large model is used to receive the problem to be processed by the user input, and to reason about the problem to be processed based on the jump gating module, the loop gating module and the optimal dynamic decision rule. During the reasoning process, it dynamically expands its multiple network layers and obtains the output result based on the expanded multiple network layers.

[0073] It should be noted that, Figure 1 The core of the technical solution presented lies in receiving a set of sample questions (training dataset) input by the user and systematically training a large model using these sample questions to generate an optimized, trained large model. This large model integrates a skip gating module and a loop gating module, which play crucial roles in the training process. These modules enable the large model to learn from the sample question set and gradually explore and experiment to ultimately master the optimal dynamic decision rule. The optimal dynamic decision rule is used to determine which network layers to skip and which to execute repeatedly based on the problem to be addressed. After training, the large model can receive user-input questions and, relying on the skip gating module, the loop gating module, and the learned optimal dynamic decision rule, perform in-depth reasoning on the problems to be addressed.

[0074] Furthermore, the skip gating module and loop gating module in this invention employ the softmax function to achieve intelligent decision-making. When processing the user-input question, the skip gating module first scores each network layer and performs softmax normalization on the scores, transforming the original scores into a probability distribution between 0 and 1, ultimately generating a skip index. For example, the optimal dynamic decision rule could be to sort the probabilities of the 10 network layers from high to low, determining the first 8 network layers as skippable layers and the last two as layers to be activated; or, setting a threshold of 0.2, with network layers having a probability ≥ 0.2 being skippable layers and network layers having a probability < 0.2 being activated layers; or, with network layers having a probability ≥ 0.2 being activated layers and network layers having a probability < 0.2 being skippable layers.

[0075] During inference, the large model possesses the ability to dynamically expand multiple network layers. It can flexibly adjust its network structure according to the complexity and requirements of the specific problem, achieving more efficient computation and more accurate results. Specifically, the optimal dynamic decision rule guides the large model to identify which network layers can be skipped and which need to be executed repeatedly during inference, based on the specific problem to be addressed. This ensures inference quality while minimizing unnecessary computation, saving computing resources, and avoiding waste.

[0076] In one possible implementation, step S2 includes:

[0077] Step S21: Based on the sample problem set, train the large model with a preset loss function, aiming to minimize the loss function, and obtain the trained large model;

[0078] The loss function is used to measure the difference between the output of the large model and the standard answer, and / or to measure the use of computing resources during the training of the large model.

[0079] It should be noted that during the training phase, this invention uses a user-provided set of sample questions to optimize and train a large model with pre-set jump gating and loop gating modules. The training process is guided by a preset loss function, which simultaneously monitors two key metrics: the content difference between the large model's output and the standard answer, and the usage of computing resources during training (such as peak GPU memory usage and training time per session). By continuously adjusting the parameters of the large model (such as the parameters of the jump gating module, the loop gating module, and the network layers), the loss function is minimized, ultimately resulting in a trained large model that can accurately answer questions while intelligently controlling computing resource consumption.

[0080] To facilitate understanding, the training process of a large model can be compared to a student reviewing for an exam. In the initial review phase, the student randomly selects review content, potentially spending a lot of time on easy questions (similar to the large model's random attempts during training, failing to skip layers it should have), or skipping key knowledge points (similar to the large model incorrectly skipping crucial layers during training). The result is a poor exam score (similar to the large model making incorrect predictions). Similarly, when grading the exam, the teacher can provide two types of feedback (similar to a loss function): more points are deducted for incorrect answers (similar to being penalized for incorrect predictions); fewer points are deducted for correct but lengthy answers (similar to being penalized for correct predictions but excessive computational resource consumption). The student can adjust their review strategy accordingly, discovering in the next review that certain question types require detailed calculations or have fixed answer formulas (the process of adjusting review strategies is similar to the parameter update process of a large model). Subsequently, students will learn to perform detailed calculations when encountering certain complex problem types, and to directly apply formulas when encountering certain simple problem types (similar to how a large model gradually stabilizes in the later stages of training and learns the optimal dynamic decision rules).

[0081] In one possible implementation, the architecture of the trained large model is a Transformer architecture, which includes at least an embedding layer and multiple network layers.

[0082] The jump gate module is used to receive the output information of the embedding layer, and output the jump index based on the output information of the embedding layer and the optimal dynamic decision rule. The jump index is used to indicate the jump execution state of each network layer in multiple network layers. The jump execution state includes skipping the current network layer and activating the current network layer.

[0083] The loop gating module is used to receive the output information of the embedding layer, and output the loop index based on the output information of the embedding layer and the optimal dynamic decision rule. The loop index is used to indicate the loop information of each network layer in multiple network layers.

[0084] The loop information includes any of the following: the current network layer belongs to a loop group and the number of loops in the loop group; the current network layer does not belong to a loop group. Wherein, if the current network layer is selected independently or not selected, the loop information corresponding to the current network layer is: the current network layer does not belong to a loop group; if the current network layer is selected, and at least one of the network layers adjacent to the current network layer is also selected, the loop information corresponding to the current network layer is: the current network layer belongs to a loop group and the number of loops in the loop group.

[0085] Multiple network layers are used to determine whether to jump based on the jump index and whether to loop based on the loop index.

[0086] It should be noted that this can be used as a reference. Figure 2 Existing large models that use the Transformer architecture have at least an embedding layer, multiple network layers (Layer 1, Layer 2, Layer N), and a language model head (LM Head). It can be seen that its network structure is a linear static structure, and the number of its network layers is fixed during the training and use of large models, which leads to large models consuming a lot of computing resources during inference and training.

[0087] This invention improves upon this approach; the improved network hierarchy structure diagram of the large model can be found by reference. Figure 3 , Figure 3 The large model shown, employing the Transformer architecture, includes at least: an embedding layer, multiple network layers (Layer 1, Layer i, Layer i+1, Layer j, Layer j+1, Layer j+2, Layer N), a language model head (LM Head), a skip gate, and a recurrent gate. First, the embedding layer transforms the input problem into a vector representation suitable for the large model, providing foundational information for subsequent network layers. The skip gate receives the output from the embedding layer and generates a skip index based on the optimal dynamic decision rule. The skip index indicates the skip execution state of each network layer (i.e., if the number of network layers is N, the dimension of the skip index is also N, where N is a positive integer), including whether the current network layer is skipped or activated, such as... Figure 3 In the skip index, blue squares indicate that the current network layer is activated, and black squares indicate that the current network layer is skipped. In this way, large models can effectively determine which layers are necessary and which can be skipped during inference, based on the problem to be addressed, thereby avoiding unnecessary computation and saving computing resources while ensuring the performance of large models.

[0088] Meanwhile, the recurrent gating module also receives the output information from the embedded layer and generates a recurrent index based on the optimal dynamic decision rule. The recurrent index indicates the recurrent information of each network layer among multiple network layers. This recurrent information can be whether the current network layer belongs to a recurrent group and the number of iterations in that group, or whether the current network layer does not belong to a recurrent group. Specifically, if the current network layer is selected independently (e.g., ...), ... Figure 3(The isolated orange square in the recurrent index) or not selected (such as...) Figure 3 If the black square represents the recurrent index, its loop information is "not part of the loop group"; however, if the current network layer is selected and at least one of its adjacent network layers is also selected (e.g., ...), then its loop information is "not part of the loop group". Figure 3 If three consecutive orange squares exist in the recurrent index, then its cycle information is "belongs to the cycle group" along with the cycle count.

[0089] Furthermore, during the execution of multiple network layers, the large model will decide whether to skip certain network layers based on the skip index, and determine whether to perform cyclic execution of network layers based on the loop index.

[0090] In summary, multiple network layers automatically skip unnecessary computations based on the skip index, while repeating computations a specified number of times on network layers marked as cyclic groups based on the cycle index. This allows large models to flexibly adjust their network structure to adapt to different input requirements and computing resources when dealing with problems, thereby saving computing resources while ensuring the performance of large models.

[0091] In one possible implementation, the number of loops is related to the complexity of the problem to be processed; the higher the complexity of the problem, the greater the number of loops.

[0092] It's important to note that in practice, the loop gating module dynamically adjusts the number of loops based on the complexity of the problem. When dealing with simple problems, it automatically detects low-complexity features (such as short texts or common word combinations) and sets the loop count of the relevant network layers to a basic computation of 1. For high-complexity problems, it increases the loop count of the core network layers by analyzing the structure of long texts, the density of technical terms, and the depth of logical nesting. Thus, by dynamically adjusting the loop count to adapt to the complexity of the problem, the large model can perform basic computations on simple problems, while increasing the loop count of the core network layers on high-complexity problems, thereby saving computational resources while optimizing inference efficiency and accuracy.

[0093] It should also be noted that problems can be labeled during the training phase of large models. For example, problems can be divided into simple problem sample groups and complex problem sample groups. For complex problems, a higher number of iterations can be preset, while for simple problems, a lower number of iterations can be preset, or no iterations can be performed. In this way, manual intervention can be used to save computing resources while ensuring the performance of large models.

[0094] It should also be noted that the number of iterations can be determined in two ways: either by manually setting a fixed value (such as setting all iteration groups to execute 3 times), or by dynamically adjusting the number of iterations based on the Poisson distribution.

[0095] In one possible implementation, if there is a skip network layer in the loop group that is in the skip execution state, then the difference between the total number of network layers in the loop group and the number of skip network layers is determined. If the difference is greater than or equal to 2, then the remaining network layers in the loop group, excluding the skip network layers, are looped.

[0096] It should be noted that when there is a skipped network layer in the loop group whose execution status is "skip current network layer", the difference between the total number of network layers in the loop group and the number of skipped network layers must first be calculated. If the difference is ≥ 2 (for example, if the loop group contains 5 network layers, of which 2 are marked as skipped, then there are 3 remaining, 3 ≥ 2), then the remaining 3 network layers that were not skipped are subjected to a specified number of loop calculations. Or as follows: Figure 3 As shown, Layer j, Layer j+1, and Layer j+2 form a loop group. However, the skip execution state of Layer j+1 is "skip the current network layer." Since the difference between the total number of network layers in this loop group and the number of skipped network layers is 3-1=2, meaning the difference 2 satisfies the condition ≥2, the remaining two network layers, Layer j and Layer j+2, are looped. Therefore, the large model can effectively identify and perform loop calculations on the network layers within the loop group that have not been skipped, thus ensuring that the computational power of the network layers within the loop group can still be fully utilized during the reasoning process of complex problems, optimizing reasoning efficiency and improving the accuracy of the reasoning results.

[0097] In summary, the large model incorporates a skip-gating module and a loop-gating module. During training, the large model learns from a sample problem set, gradually acquiring the optimal dynamic decision rule. The skip-gating module allows the large model to intelligently skip unnecessary network layers during inference, based on the problem to be addressed and the optimal dynamic decision rule, thus reducing computational load and resource consumption. The loop-gating module, on the other hand, can repeatedly execute certain key network layers when needed (e.g., when the problem is complex), enhancing the large model's understanding and handling of complex problems. This ability to dynamically expand the number of network layers in the large model enables the trained model to adaptively adjust network layers according to the problem to be addressed (the actual input), thereby ensuring inference performance while saving computational resources, avoiding waste, and improving resource utilization.

[0098] Figure 4 This invention illustrates a large-scale network dynamic expansion device for an intelligent computing center cloud platform that provides computing power resources, as provided by the present invention. Figure 4 As shown, the device 40 includes:

[0099] The receiving module 401 is used to perform step S1: receiving a set of sample questions input by the user;

[0100] Execution module 402 is used to execute step S2: training the large model based on the sample problem set to obtain the trained large model; wherein, the large model is equipped with a jump gate module and a loop gate module;

[0101] The jump gating module and the loop gating module are used to learn the optimal dynamic decision rule gradually based on the sample problem set during the training of a large model.

[0102] The trained large model is used to receive user input of the problem to be processed, and to reason about the problem based on the jump gating module, the loop gating module and the optimal dynamic decision rule. During the reasoning process, it dynamically expands its multiple network layers and obtains the output result based on the expanded multiple network layers.

[0103] Among them, the optimal dynamic decision rule is used to determine the network layers that need to be skipped and the network layers that need to be executed repeatedly among multiple network layers based on the problem to be solved.

[0104] In one possible implementation, step S2 includes:

[0105] Step S21: Based on the sample problem set, train the large model with a preset loss function, aiming to minimize the loss function, and obtain the trained large model;

[0106] The loss function is used to measure the difference between the output of the large model and the standard answer, and / or to measure the use of computing resources during the training of the large model.

[0107] In one possible implementation, the architecture of the trained large model is a Transformer architecture. The trained large model with a Transformer architecture includes at least an embedding layer and multiple network layers.

[0108] The jump gate module is used to receive the output information of the embedding layer, and output the jump index based on the output information of the embedding layer and the optimal dynamic decision rule. The jump index is used to indicate the jump execution state of each network layer in multiple network layers. The jump execution state includes skipping the current network layer and activating the current network layer.

[0109] The loop gating module is used to receive the output information of the embedding layer, and output the loop index based on the output information of the embedding layer and the optimal dynamic decision rule. The loop index is used to indicate the loop information of each network layer in multiple network layers.

[0110] Cyclic information includes any of the following: the current network layer belongs to a cyclic group and the number of cycles in the cyclic group; the current network layer does not belong to a cyclic group.

[0111] If the current network layer is selected independently or not selected, the corresponding loop information for the current network layer is: the current network layer does not belong to the loop group; if the current network layer is selected, and at least one of the network layers adjacent to the current network layer is also selected, the corresponding loop information for the current network layer is: the current network layer belongs to the loop group and the number of loops in the loop group;

[0112] Multiple network layers are used to determine whether to jump based on the jump index and whether to loop based on the loop index.

[0113] In one possible implementation, the number of loops is related to the complexity of the problem to be processed; the higher the complexity of the problem, the greater the number of loops.

[0114] In one possible implementation, if there is a skip network layer in the loop group that is in the skip execution state, then the difference between the total number of network layers in the loop group and the number of skip network layers is determined. If the difference is greater than or equal to 2, then the remaining network layers in the loop group, excluding the skip network layers, are looped.

[0115] In summary, the large model incorporates a skip-gating module and a loop-gating module. During training, the large model learns from a sample problem set, gradually acquiring the optimal dynamic decision rule. The skip-gating module allows the large model to intelligently skip unnecessary network layers during inference, based on the problem to be addressed and the optimal dynamic decision rule, thus reducing computational load and resource consumption. The loop-gating module, on the other hand, can repeatedly execute certain key network layers when needed (e.g., when the problem is complex), enhancing the large model's understanding and handling of complex problems. This ability to dynamically expand the number of network layers in the large model enables the trained model to adaptively adjust network layers according to the problem to be addressed (the actual input), thereby ensuring inference performance while saving computational resources, avoiding waste, and improving resource utilization.

[0116] Please refer to Figure 5 The present invention also provides an electronic device 50, including a processor 501, a memory 502, and a computer program stored in the memory 502 and executable on the processor 501. When the computer program is executed by the processor 501, it implements the steps of the above-mentioned method for dynamic expansion of the large model network of the intelligent computing center cloud platform that provides computing power resources, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0117] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the above-described method for dynamically expanding the large-scale network of an intelligent computing center cloud platform that provides computing resources, and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0118] The present invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the above-described method for dynamically expanding the large model network of the intelligent computing center cloud platform that provides computing power resources, and can achieve the same technical effect. To avoid repetition, the details will not be repeated here.

[0119] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that the above methods can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the present invention.

[0121] The present invention has been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit and scope of the claims, and all such modifications are within the protection scope of the present invention.

Claims

1. A method for dynamically expanding a large-scale model network of an intelligent computing center cloud platform that provides computing power resources, characterized in that, The method includes: Step S1: Receive the sample question set input by the user; Step S2: Based on the sample problem set, train the large model to obtain the trained large model; wherein, the large model is equipped with a jump gate module and a loop gate module; The jump gate module and the loop gate module are used to learn based on the sample problem set during the training process of the large model, so as to gradually learn the optimal dynamic decision rule. The trained large model is used to receive the user input of the problem to be processed, and to reason about the problem based on the jump gating module, the loop gating module and the optimal dynamic decision rule. During the reasoning process, it dynamically expands its multiple network layers and obtains the output result based on the expanded multiple network layers. The optimal dynamic decision rule is used to determine, based on the problem to be solved, the network layers that need to be skipped and the network layers that need to be executed cyclically among the multiple network layers; Step S2 includes: Step S21: Based on the sample problem set, train the large model using a preset loss function, with the goal of minimizing the loss function, to obtain the trained large model; The loss function is used to measure the difference between the output of the large model and the standard answer, and / or to measure the use of computing resources during the training of the large model. The trained large model has a Transformer architecture, and the trained large model with the Transformer architecture includes at least an embedding layer and the multiple network layers. The jump gate module is used to receive the output information of the embedding layer, and output a jump index based on the output information of the embedding layer and the optimal dynamic decision rule. The jump index is used to indicate the jump execution state of each of the plurality of network layers. The jump execution state includes any one of the following: skip the current network layer, activate the current network layer. The loop gating module is used to receive the output information of the embedding layer, and output a loop index based on the output information of the embedding layer and the optimal dynamic decision rule, wherein the loop index is used to indicate the loop information of each network layer in the plurality of network layers; The cycle information includes any one of the following: the current network layer belongs to a cycle group and the number of cycles in the cycle group; the current network layer does not belong to a cycle group. Wherein, if the current network layer is selected independently or not selected, the loop information corresponding to the current network layer is: the current network layer does not belong to the loop group; if the current network layer is selected, and at least one of the network layers adjacent to the current network layer is also selected, the loop information corresponding to the current network layer is: the current network layer belongs to the loop group and the number of loops of the loop group; The plurality of network layers are used to determine whether to perform a jump based on the jump index and whether to perform a loop based on the loop index.

2. The method according to claim 1, characterized in that, The number of iterations is related to the complexity of the problem to be processed; the higher the complexity of the problem, the greater the number of iterations.

3. The method according to claim 1, characterized in that, If there is a skip network layer in the loop group that skips the current network layer, then the difference between the total number of network layers in the loop group and the number of skip network layers is determined. If the difference is greater than or equal to 2, then the remaining network layers in the loop group, except for the skip network layer, are looped.

4. A dynamic expansion device for a large-scale model network of an intelligent computing center cloud platform that provides computing power resources, characterized in that: The device includes: The receiving module is used to perform step S1: receiving a set of sample questions input by the user; The execution module is used to execute step S2: based on the sample problem set, train the large model to obtain the trained large model; wherein, the large model is equipped with a jump gate module and a loop gate module; The jump gate module and the loop gate module are used to learn based on the sample problem set during the training process of the large model, so as to gradually learn the optimal dynamic decision rule. The trained large model is used to receive the user input of the problem to be processed, and to reason about the problem based on the jump gating module, the loop gating module and the optimal dynamic decision rule. During the reasoning process, it dynamically expands its multiple network layers and obtains the output result based on the expanded multiple network layers. The optimal dynamic decision rule is used to determine, based on the problem to be solved, the network layers that need to be skipped and the network layers that need to be executed cyclically among the multiple network layers; Step S2 includes: Step S21: Based on the sample problem set, train the large model using a preset loss function, with the goal of minimizing the loss function, to obtain the trained large model; The loss function is used to measure the difference between the output of the large model and the standard answer, and / or to measure the use of computing resources during the training of the large model. The trained large model has a Transformer architecture, and the trained large model with the Transformer architecture includes at least an embedding layer and the multiple network layers. The jump gate module is used to receive the output information of the embedding layer, and output a jump index based on the output information of the embedding layer and the optimal dynamic decision rule. The jump index is used to indicate the jump execution state of each of the plurality of network layers. The jump execution state includes any one of the following: skip the current network layer, activate the current network layer. The loop gating module is used to receive the output information of the embedding layer, and output a loop index based on the output information of the embedding layer and the optimal dynamic decision rule, wherein the loop index is used to indicate the loop information of each network layer in the plurality of network layers; The cycle information includes any one of the following: the current network layer belongs to a cycle group and the number of cycles in the cycle group; the current network layer does not belong to a cycle group. Wherein, if the current network layer is selected independently or not selected, the loop information corresponding to the current network layer is: the current network layer does not belong to the loop group; if the current network layer is selected, and at least one of the network layers adjacent to the current network layer is also selected, the loop information corresponding to the current network layer is: the current network layer belongs to the loop group and the number of loops of the loop group; The plurality of network layers are used to determine whether to perform a jump based on the jump index and whether to perform a loop based on the loop index.

5. A server, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of a method for dynamically expanding a large-scale network of an intelligent computing center cloud platform that provides computing power resources as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for dynamically expanding a large-scale network of an intelligent computing center cloud platform that provides computing power resources as described in any one of claims 1-3.

7. A computer program product, characterized in that, The method includes computer instructions that, when executed by a processor, implement the steps of a method for dynamically expanding a large-scale network of an intelligent computing center cloud platform that provides computing power resources as described in any one of claims 1-3.

Images