A model fine-tuning method based on an intelligent computing platform and an intelligent computing platform

Abstract

The application discloses a model fine-tuning method and a wisdom computing platform based on the wisdom computing platform, the wisdom computing platform comprises a user end, a dispatching center and a computing cluster, the method is applied to the dispatching center side and comprises the following steps: receiving a model fine-tuning task sent by the user end, the model fine-tuning task at least comprises a bid of a user, a fine-tuning data set and a computing demand quantity; based on the bid, the fine-tuning data set and the computing demand quantity, in combination with node capability information of each computing node in the computing cluster, an optimal scheduling combination of the model fine-tuning task is determined with the maximization of total income as a target, wherein the optimal scheduling combination comprises a target computing node; whether the model fine-tuning task is accepted or not is judged according to the optimal scheduling combination; if the model fine-tuning task is accepted, the target computing node is controlled to execute the model fine-tuning task based on the fine-tuning data set. The application can improve the resource utilization rate of the computing node and provide a flexible pricing scheme to adapt to the changing market supply and demand relationship.

Description

Technical Field

[0001] This invention relates to the technical field of large model fine-tuning, and in particular to a model fine-tuning method and intelligent computing platform based on an intelligent computing platform. Background Technology

[0002] Large-scale AI models refer to deep learning models with a large number of parameters trained using massive amounts of data and powerful computing capabilities. These models typically possess high versatility and generalization ability, and can be applied to intelligent scenarios such as natural language processing, image recognition, and speech recognition. However, large-scale models do not perform well in some niche domains or specific scenarios. Therefore, it is necessary to fine-tune general-purpose large-scale models on large amounts of specialized data to make them suitable for certain specific scenarios.

[0003] In related technologies, large model fine-tuning tasks are essentially deep learning training tasks. However, the scheduling of deep learning training tasks has not fully considered the key characteristics of large model fine-tuning tasks, such as computational and memory capabilities, leading to low resource utilization. Furthermore, existing pricing schemes for large model fine-tuning tasks have failed to adapt to the ever-changing supply and demand dynamics of the market. Summary of the Invention

[0004] In view of this, the present invention provides a model fine-tuning method and an intelligent computing platform based on an intelligent computing platform, which can improve the resource utilization of computing nodes and provide a flexible pricing scheme to adapt to the ever-changing market supply and demand relationship.

[0005] A first aspect of the present invention provides a model fine-tuning method based on an intelligent computing platform, the intelligent computing platform including a user terminal, a scheduling center, and a computing cluster, the method being applied to the scheduling center side, the method comprising:

[0006] Receive the model fine-tuning task sent by the user terminal, wherein the model fine-tuning task includes at least the user's bid, the fine-tuning dataset, and the computational requirements;

[0007] Based on the bid, the fine-tuning dataset, and the computational requirements, and combined with the node capability information of each computing node in the computing cluster, the optimal scheduling combination for the model fine-tuning task is determined with the goal of maximizing total revenue. The optimal scheduling combination includes target computing nodes selected from the computing cluster that meet the computational requirements. The node capability information includes computing power, video memory capacity, and video memory bandwidth capacity. The total revenue includes the user's revenue and the revenue of the intelligent computing service provider.

[0008] Determine whether to accept the model fine-tuning task based on the optimal scheduling combination;

[0009] If the model fine-tuning task is accepted, the target computing node is controlled to execute the model fine-tuning task based on the fine-tuning dataset.

[0010] A second aspect of the present invention provides an intelligent computing platform, the intelligent computing platform including a user terminal, a scheduling center, and a computing cluster, wherein the scheduling center includes the following modules:

[0011] The task receiving module is used to receive the model fine-tuning task sent by the user terminal. The model fine-tuning task includes the user's bid, the fine-tuning dataset, and the computational requirements.

[0012] The optimal scheduling combination determination module is used to determine the optimal scheduling combination for the model fine-tuning task based on the bid, the fine-tuning dataset, and the computing requirements, combined with the node capability information of each computing node in the computing cluster, with the objective of maximizing total revenue. The optimal scheduling combination includes target computing nodes selected from the computing cluster that meet the computing requirements; the node capability information includes computing power, video memory capacity, and video memory bandwidth capacity; and the total revenue includes the user's revenue and the revenue of the intelligent computing service provider.

[0013] The task acceptance judgment module is used to determine whether to accept the model fine-tuning task based on the optimal scheduling combination.

[0014] The task execution module is used to control the target computing node to execute the model fine-tuning task based on the fine-tuning dataset if the model fine-tuning task is accepted.

[0015] A third aspect of the present invention provides an electronic device, the electronic device comprising:

[0016] At least one processor; and

[0017] A memory communicatively connected to the at least one processor; wherein,

[0018] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the model fine-tuning method based on the intelligent computing platform as described in the first aspect above.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the model fine-tuning method based on an intelligent computing platform as described in the first aspect above.

[0020] The fifth aspect of the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the model fine-tuning method based on an intelligent computing platform as described in the first aspect above.

[0021] In this embodiment, the intelligent computing platform enables unified scheduling and management of model fine-tuning tasks from different user terminals, which facilitates flexible and precise control of the tasks. Each model fine-tuning task includes a fine-tuning dataset and computational requirements. When determining the target computing node from the computing cluster, the platform considers node capability information such as computing power, memory capacity, and memory bandwidth. The selected target computing node must meet the computational requirements, which is beneficial for the smooth execution of the task and for selecting a matching computing node, thereby improving the resource utilization of the computing nodes.

[0022] Furthermore, the model fine-tuning task in this embodiment includes the user's bid. When determining whether to accept the task, the user's bid also needs to be considered. With the goal of maximizing total revenue, the optimal scheduling combination of the model fine-tuning task is determined, and the actual payment fee is calculated based on the optimal scheduling combination. By designing a pricing scheme for users and intelligent computing service providers through this auction mechanism, the total revenue of users and service providers can be maximized, while flexibly adapting to the ever-changing market supply and demand relationship.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the framework of an intelligent computing platform provided in Embodiment 1 of the present invention.

[0026] Figure 2 This is a flowchart of a model fine-tuning method based on an intelligent computing platform provided in Embodiment 1 of the present invention.

[0027] Figure 3 This is a schematic diagram of a multi-LoRA task training method provided in Embodiment 1 of the present invention;

[0028] Figure 4This is a flowchart of a model fine-tuning method based on an intelligent computing platform provided in Embodiment 2 of the present invention;

[0029] Figure 5 This is a flowchart of a model fine-tuning method based on an intelligent computing platform provided in Embodiment 3 of the present invention;

[0030] Figure 6 This is a schematic diagram of the structure of a scheduling center provided in Embodiment 4 of the present invention.

[0031] Figure 7 This is a schematic diagram of the structure of an electronic device provided in Embodiment 5 of the present invention. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be used interchangeably where appropriate so that the embodiments of the invention described herein can cover implementations in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] Example 1

[0035] Embodiment 1 of this invention provides a model fine-tuning method based on an intelligent computing platform. This method can be executed by the intelligent computing platform, which can be implemented in hardware and / or software. See also Figure 1 This diagram illustrates the framework of an intelligent computing platform, a cloud service platform that provides services to intelligent computing service providers, including a user terminal, a scheduling center, and a computing cluster.

[0036] The user-side is the client or terminal for user interaction, providing an interactive interface for users to submit model fine-tuning tasks. In implementation, large model fine-tuning often uses lightweight fine-tuning techniques such as LoRA. LoRA can keep the pre-trained model parameters unchanged and only fine-tune the parameters of the LoRA adapter. Therefore, the model fine-tuning task in this embodiment can also be called the LoRA fine-tuning task.

[0037] The user terminal connects to the scheduling center, which is used to schedule and process model fine-tuning tasks.

[0038] Connected to the scheduling center is a computing cluster, which can be a GPU cluster consisting of a group of [K] = {1,2,…,K} GPU computing nodes, used to execute user-submitted LoRA fine-tuning tasks.

[0039] In a further embodiment, such as Figure 1 As shown, connected to the scheduling center are data providers from third-party data marketplaces. These providers perform data preprocessing for LoRA fine-tuning tasks, such as labeling and cleaning. Specifically, data providers are preprocessing service providers that preprocess tasks before they are processed by the computing nodes. Users need to submit a dataset for fine-tuning each task, but current large-scale model fine-tuning has strict requirements on the dataset format. For ordinary users, directly submitting datasets that meet the format specified by the intelligent computing platform remains a significant challenge; therefore, third-party data marketplaces can be used to assist with data preprocessing. There can be multiple data providers, which can be represented in this embodiment as [N] = {1, 2, ..., N}.

[0040] based on Figure 1 The intelligent computing platform, see Figure 2 The flowchart shown is a model fine-tuning method based on an intelligent computing platform provided in Embodiment 1 of the present invention. This embodiment can be executed by a scheduling center, such as... Figure 2 As shown, this embodiment may include the following steps:

[0041] Step 101: Receive the model fine-tuning task sent by the user. The model fine-tuning task includes at least the user's bid, the fine-tuning dataset, and the computational requirements.

[0042] In implementation, the scheduling center can form the received model fine-tuning tasks into a fine-tuning task set [I] = {1,2,...,I}. It should be noted that this embodiment can be applied to scenarios where fine-tuning is performed on a single large model, in which case the fine-tuning task set targets the same large model; it can also be applied to scenarios where fine-tuning is performed on multiple large models, in which case different large models can correspond to different fine-tuning task sets.

[0043] Each model fine-tuning task i (hereinafter referred to as task i) may include the following information: task arrival time, task deadline, training dataset used for fine-tuning (i.e., fine-tuning dataset), GPU memory requirements (i.e., GPU memory requirements), computational requirements (i.e., total computational cost required), whether data preprocessing is required, user bids, etc.

[0044] As an example, the model fine-tuning task i can be represented as Among them, a i Indicates the task arrival time; d i This indicates the task deadline, i.e., task i needs to be completed by d. i This must be completed beforehand; otherwise, the computing service provider will not receive payment for performing the task. This represents the training dataset used for fine-tuning (i.e., the fine-tuning dataset); r i M represents the video memory requirement of task i; i This represents the computational requirement of task i, i.e., the total number of data samples that need to be processed; f i Indicates whether task i requires data preprocessing, b i This represents the user's bid, i.e., if the intelligent computing service provider meets the user's specified deadline d. i The fee the user is willing to pay upon completion of task i.

[0045] In this embodiment, users can submit model fine-tuning tasks through a bidding process on the user client, facilitating the submission of personalized fine-tuning requests. The model fine-tuning task includes the fine-tuning dataset and computational requirements, which helps in selecting suitable computing nodes for task processing during subsequent computing node matching, thereby improving the resource utilization of computing nodes. The model fine-tuning task also includes the user's bid, allowing users to better express their bidding intentions without being limited by a fixed price, thus increasing pricing flexibility.

[0046] Step 102: Based on the bid, the fine-tuning dataset, and the computational requirements, and combined with the node capability information of each computing node in the computing cluster, determine the optimal scheduling combination of the model fine-tuning tasks with the goal of maximizing total revenue. The optimal scheduling combination includes target computing nodes selected from the computing cluster that meet the computational requirements.

[0047] In this step, after the model fine-tuning task arrives at the intelligent computing platform, the scheduling center determines the optimal scheduling combination of the model fine-tuning task based on the demand information in the model fine-tuning task and the node capability information of each computing node in the computing cluster, with the goal of maximizing the total benefit.

[0048] Maximizing total revenue can be understood as maximizing social welfare. In this embodiment, total revenue includes the user's revenue and the revenue of the intelligent computing service provider. In one implementation, the user's revenue can be the cost savings achieved by the pricing scheme in this embodiment. The intelligent computing service provider's revenue can be the profit remaining after deducting all expenses from the fees received from the user.

[0049] In this embodiment, when determining the target computing node, the node capability information of each computing node in the computing cluster is considered. For example, the node capability information may include computing power, video memory capacity, and video memory bandwidth capacity. Computing power refers to the indicator that measures the processing ability and performance of a computing node, representing the maximum number of data samples that the computing node can process in a single time period. Video memory capacity refers to the video memory capacity. Video memory is the storage device in the graphics card used to store image data. Its capacity determines the number and resolution of images that the graphics card can process. The larger the video memory, the more images can be stored. The unit is GB. Video memory bandwidth refers to the data transfer rate between the display chip and the video memory. The unit is GB / s. The larger the video memory bandwidth, the faster the GPU processes data. Video memory bandwidth is determined by both the video memory frequency and the video memory bus width.

[0050] Scheduling composition refers to the combination of various execution elements that satisfy the requirements of a model fine-tuning task. These execution elements may include, but are not limited to: task acceptance information (i.e., whether the current task i has been accepted), computing node information, and data provider information. For example, scheduling composition can be represented as l = {u i ,{x ikt} k,t ,{z in} n}, where l represents scheduling combination, u i Indicates whether the current task i has been accepted; {x ikt} k,t This indicates the information about the computing node, specifically whether task i was executed on computing node k during time period t; {z in} n This indicates the data provider information, specifically whether or not the services of data provider n are selected and used for data preprocessing of task i.

[0051] Step 103: Determine whether to accept the model fine-tuning task based on the optimal scheduling combination.

[0052] In this step, u can be determined based on the optimal scheduling combination. i The value of u determines whether to accept the model fine-tuning task. For example, if u i =0, then it is determined that the model fine-tuning task is not accepted; if u i If the value is greater than 0, it is determined that the task of fine-tuning the model is accepted.

[0053] Step 104: If the model fine-tuning task is accepted, control the target computing node to execute the model fine-tuning task based on the fine-tuning dataset.

[0054] In this step, when task i is received, it can be sent to the target computing node. The target computing node can then use LoRA fine-tuning technology based on the training dataset in task i to fine-tune the large model before the task deadline. The platform can be configured to run in discrete time intervals [T] = {1, 2, ..., T}.

[0055] In one implementation, when the target compute node performs model fine-tuning tasks, it can train a large model based on shared pre-trained parameters. For example, as Figure 3 As shown in the diagram of multi-LoRA task training, LoRA can keep the pre-trained model parameters unchanged and only fine-tune the parameters of the LoRA adapter. Here, (x1, x2, x3) are the training data points for the three fine-tuning tasks. During the forward propagation process, the three tasks calculate the forward output based on the pre-trained parameters and the corresponding LoRA adapter, further calculate the loss function, and backpropagate the gradient to the corresponding LoRA adapter for parameter updates. Sharing pre-trained parameters among different LoRA fine-tuning tasks can further reduce the resources required for the fine-tuning process.

[0056] In one embodiment, if task i chooses to use a data provider service for data preprocessing, the scheduling center selects a target data provider from a third-party data marketplace before sending task i to the target computing node, and sends the training dataset of task i to the target data provider for data preprocessing. After obtaining the preprocessed training dataset from the target data provider, the scheduling center replaces the training dataset in task i with the preprocessed training dataset and sends task i to the target computing node.

[0057] When selecting a target data provider from a third-party data marketplace, the intelligent computing platform will comprehensively consider the price q charged by each data provider n. in and the latency h required for each data provider n to process the training dataset for task i. in Then, select one of the data providers' services for task i.

[0058] In this embodiment, the intelligent computing platform enables unified scheduling and management of model fine-tuning tasks from different user terminals, which facilitates flexible and precise control of the tasks. Each model fine-tuning task includes a fine-tuning dataset and computational requirements. When determining the target computing node from the computing cluster, the platform considers node capability information such as computing power, memory capacity, and memory bandwidth. The selected target computing node must meet the computational requirements, which is beneficial for the smooth execution of the task and for selecting a matching computing node, thereby improving the resource utilization of the computing nodes.

[0059] Example 2

[0060] See Figure 4 This document illustrates a flowchart of a model fine-tuning method based on an intelligent computing platform, as provided in Embodiment 2 of the present invention. This embodiment, based on Embodiment 1, explains the pricing process for tasks, as follows: Figure 4 As shown, this embodiment may include the following steps:

[0061] Step 201: Receive the model fine-tuning task sent by the user. The model fine-tuning task includes at least the user's bid, the fine-tuning dataset, and the computational requirements.

[0062] Step 202: Based on the bid, the fine-tuning dataset, and the computational requirements, and combined with the node capability information of each computing node in the computing cluster, determine the optimal scheduling combination of the model fine-tuning tasks with the goal of maximizing the total revenue. The optimal scheduling combination includes target computing nodes selected from the computing cluster that meet the computational requirements.

[0063] For example, node capability information includes computing power, video memory capacity, and video memory bandwidth capacity; total revenue includes revenue for users and revenue for intelligent computing service providers.

[0064] Step 203: Determine whether to accept the model fine-tuning task based on the optimal scheduling combination.

[0065] Step 204: If the model fine-tuning task is accepted, control the target computing node to execute the model fine-tuning task based on the fine-tuning dataset.

[0066] Step 205: Calculate the actual payment cost based on the optimal scheduling combination.

[0067] In this embodiment, the actual payment may differ from the user's bid; generally, the actual payment is lower than the user's bid. If the actual payment is higher than the user's bid, the task is rejected, and the user does not need to pay.

[0068] In practice, the actual payment can be the expenses incurred by the platform for executing the task. These expenses include fees paid to third-party data providers, computing costs of computing nodes, memory costs, and memory bandwidth costs.

[0069] Step 206: Return the actual payment amount to the user so that the user can make payment according to the actual payment amount.

[0070] In this embodiment, the model fine-tuning task includes the user's bid. When deciding whether to accept the task, the user's bid also needs to be considered. With the goal of maximizing total revenue, the optimal scheduling combination of the model fine-tuning task is determined, and the actual payment fee is calculated based on the optimal scheduling combination. By designing a pricing scheme for users and intelligent computing service providers through this auction mechanism, the total revenue of users and service providers can be maximized, while flexibly adapting to the ever-changing market supply and demand relationship.

[0071] Example 3

[0072] See Figure 5 This document illustrates a flowchart of a model fine-tuning method based on an intelligent computing platform, provided in Embodiment 3 of the present invention. This embodiment, based on Embodiment 1 or Embodiment 2, describes the process of determining the optimal scheduling combination, as follows: Figure 5 As shown, this embodiment may include the following steps:

[0073] Step 301: Receive the model fine-tuning task sent by the user. The model fine-tuning task includes at least the user's bid, the fine-tuning dataset, and the computational requirements.

[0074] Step 302: Based on the bid, fine-tuning dataset, and computational requirements, and combined with the node capability information of each computing node in the computing cluster, a primitive optimization problem based on decision variables is modeled with the goal of maximizing total revenue.

[0075] Among them, the data provider is a preprocessing service provider that preprocesses tasks before they are processed on the computing nodes.

[0076] The original optimization problem includes an optimization objective and a first set of constraints. The first set of constraints includes multiple decision variables, for example, including variables related to data providers, computing nodes, and task acceptance.

[0077] In this embodiment, the optimization objective is to maximize the total revenue. In one implementation, the total revenue is represented as follows:

[0078]

[0079] in,

[0080]

[0081] Where U represents the total revenue, U r U represents the revenue earned by all users. c U represents the revenue earned by the intelligent computing service provider. i b represents the benefit gained by the user who initiates model fine-tuning task i. i p represents the bid of the user who initiated the model fine-tuning task i. i u represents the actual cost paid by the user after model fine-tuning task i is accepted. i p represents the task acceptance information indicating whether model fine-tuning task i is accepted. i u i This represents the actual payment received by the platform for model fine-tuning task i, z in This indicates whether model fine-tuning task i requires data provider n for data preprocessing information, q. in x represents the fee charged by the data provider. ikt Indicates whether model fine-tuning task i,e is executed on computing node k during time period t. ikt Let represent the system overhead caused by computing node k training model fine-tuning task i within time period t, where n represents the data provider, k represents the computing node, and t represents the execution time period of the computing node.

[0082] ∑ i ∑ n q in z in This represents the fees paid by the platform to the data provider; ∑ i ∑ k ∑ t e ikt x ikt This represents the overhead of the platform executing model fine-tuning task i.

[0083] Based on the above expression of total revenue, the primal optimization problem based on decision variables, with the goal of maximizing total revenue, can be represented as follows:

[0084]

[0085]

[0086]

[0087] Where P represents the optimization objective of maximizing total revenue, and the constraints (a1)-(a9) form the first set of constraints;

[0088] (a1) is used to ensure that for each model fine-tuning task i, if the task is accepted and requires data preprocessing, at most one data provider is selected; fi Indicates whether model fine-tuning task i requires data preprocessing;

[0089] (a2) is used to ensure that each model fine-tuning task i runs on at most one computing node k in each time period t;

[0090] (a3) is used to ensure that each model fine-tuning task i is executed only after it arrives at the system and completes data preprocessing; a i h represents the arrival time of model fine-tuning task i; in This represents the latency required by the data provider to process the training dataset for model fine-tuning task i;

[0091] (a4) is used to ensure that each model fine-tuning task i is executed before its deadline; d i This indicates the deadline for model fine-tuning task i;

[0092] (a5) is used to ensure that each model fine-tuning task i has completed sufficient computation; M i This represents the computational requirement for model fine-tuning task i; s ik This represents the amount of computation that can be completed in each time interval when model fine-tuning task i is executed on computing node k;

[0093] (a6) is used to constrain the computational capacity of each time period and each computing node; C kp This represents the computational power of a node k∈[K].

[0094] (a7) is used to constrain the memory capacity for each time period and each computing node; C km This represents the video memory capacity of computing node k;

[0095] (a8) is used to constrain the memory bandwidth capacity for each time period and each computing node; C kg This represents the memory bandwidth capability of computing node k;

[0096] (a9) specifies the range of values ​​for each decision variable. i x ikt z in All of them are decision variables.

[0097] Step 303 simplifies the original optimization problem into an equivalent problem based on scheduling combination.

[0098] Since the original optimization problem involves complex constraints, this embodiment introduces the concept of scheduling combinatorials to simplify it. Scheduling combinatorials are combinations of execution elements that satisfy the first set of constraints related to the task itself. These execution elements include at least computation node information, data provider information, and task acceptance information. In one implementation, the original optimization problem can be simplified to an equivalent problem based on scheduling combinatorials using Comp-Exp.

[0099] In implementation, for any task i, its scheduling combination l is defined as the decision variable l = {u i ,{x ikt} k,t ,{z in} n} Assignment of a set of specific values ​​that satisfy constraints (a1) to (a5).

[0100] In one embodiment, the equivalence problem based on scheduling composition can be expressed as follows:

[0101]

[0102]

[0103]

[0104] Where P1 represents maximizing the total revenue after equivalence, and constraints (b1)-(b5) form the second set of constraints; ζ i This refers to the set of all scheduling combinations for model fine-tuning task i that satisfy the necessary constraints; the necessary constraints include constraints (a1) to (a5); b il x represents the increment of the objective function value of problem P when using scheduling combination l to execute model fine-tuning task i; il Indicates whether model fine-tuning task i is executed according to scheduling combination l; s kt (il), r kt (il), σ kt (il) represent the computational cost, memory cost, and memory bandwidth cost when executing model fine-tuning task i on computing node k using scheduling combination l during time period t. b The GPU memory space occupied by pre-trained large models.

[0105] Furthermore, b il The following methods can be used to determine this:

[0106]

[0107] Among them, the decision variable z in and x iktThe value of is determined by the scheduling combination l.

[0108] s kt (il), r kt (il), σ kt (il) is determined using the following methods:

[0109] s kt (il)=s ik x ikt r kt (il)=r i x ikt , σ kt (il)=σ i x ikt

[0110] Where, x ikt ∈l. For ease of expression, t∈l is used to indicate that time period t is one of the time periods specified in the scheduling combination l, that is...

[0111]

[0112] Step 304: Based on the primal dual algorithm and the equivalent problem, generate the Lagrange dual problem corresponding to the primal optimization problem.

[0113] In this embodiment, an online scheduling algorithm is designed based on the idea of ​​online primal duality. First, the decision variable x... il The integer constraint is relaxed to x il ∈[0,1]. Next, write out the Lagrange dual problem of the primal problem P.

[0114] In one embodiment, the Lagrange dual problem is expressed as:

[0115]

[0116] Where, μ i , λ kt , γ kt Let (b1), (b2), (b3), and (b4) represent the dual variables associated with constraints (b1), (b2), (b3), and (b4), respectively.

[0117] Step 305: Determine the optimal scheduling combination of model fine-tuning tasks based on the Lagrange duality problem.

[0118] In this step, once the Lagrange dual problem is determined, the optimal scheduling combination of model fine-tuning tasks can be further determined based on the Lagrange dual problem.

[0119] In one embodiment, the Lagrange dual problem includes multiple dual variables, which include variables related to the computational overhead, memory overhead, and memory bandwidth overhead of the computing node when performing the task; then step 305 further includes the following steps:

[0120] Step 305-1: Based on the Lagrange dual problem, determine the values ​​of the dual variables corresponding to each scheduling combination.

[0121] In one implementation, the dual variable λ kt , γ kt The value can be determined in the following way:

[0122]

[0123] in, This can be understood as the increase in social welfare that can be brought about by consuming a unit of resources in a single time period. and Let λ represent the dual variable λ after the online algorithm processes task i. kt , and γ kt The value of .

[0124] In one implementation, The following method can be used for calculation:

[0125]

[0126] Step 305-2: Determine the functional representation of each scheduling combination based on the values ​​of the dual variables corresponding to each scheduling combination.

[0127] In one implementation, step 305-2 can be achieved using the following formula:

[0128]

[0129] Where F(il) represents the functional representation of scheduling combination l; (k,t)∈l refers to x in scheduling combination l ikt For a computation node k with a time period t, there are one or more (k,t) pairs in a scheduling combination l, because a task may be executed in multiple time periods.

[0130] In this case, the dual variable μ i The value can be:

[0131] μ i =max{0,F(il)}

[0132] Dual variable μ i The value of u determines the decision variable.i The value of l means that if the returned scheduling combination l i If the value of F(il) becomes negative, the system will reject task i and set μ as the negative value. i The value of μ is set to zero; conversely, if μ i If the value is greater than 0, the system accepts task i and determines the scheduling combination l. i Execute on the specified computing node and time period.

[0133] Step 305-3: Based on the functional representation of each scheduling combination, the optimal scheduling combination is determined using the maximum value independent variable point set function.

[0134] In one implementation, step 305-3 can be achieved using the following formula:

[0135]

[0136] Among them, l i This represents the optimal scheduling combination.

[0137] Step 306: Determine whether to accept the model fine-tuning task based on the optimal scheduling combination.

[0138] Step 307: If the model fine-tuning task is accepted, control the target computing node to execute the model fine-tuning task based on the fine-tuning dataset.

[0139] After the target computing node executes the model fine-tuning task, it returns the results to the user before the deadline.

[0140] Step 308: Calculate the actual payment cost based on the values ​​of the dual variables corresponding to the optimal scheduling combination and the preset pricing function.

[0141] In one implementation, the actual payment is calculated using the following formula:

[0142]

[0143] Or it can be expressed as:

[0144]

[0145] Where x ikt and z in The value is taken from the scheduling combination l; (k,t)∈l refers to x in the scheduling combination l. ikt =1, corresponding to k and t. This can be... This is considered the marginal price of computation, video memory, and video memory bandwidth after processing task i-1. A user's bid affects whether they can win the auction. If they win, the reward the user needs to pay depends on the amount of resources consumed in executing the task.

[0146] Step 309: Return the actual payment amount to the user terminal so that the user can make payment according to the actual payment amount.

[0147] This embodiment has the following beneficial effects:

[0148] (1) This embodiment considers more dimensions of resource constraints when it comes to the scheduling algorithm for large model fine-tuning tasks, including computing resources, video memory resources, and video memory bandwidth resources. By combining the computing requirements of the task itself and the fine-tuning dataset, the target computing node is determined, which fully considers the key characteristics of the large model fine-tuning task itself and the characteristics of the computing node, thereby improving the resource utilization of the computing node.

[0149] (2) To adapt to the ever-changing market supply and demand relationship, this embodiment designs a pricing scheme for users and intelligent computing service providers based on an auction mechanism. The intelligent computing service provider is the auctioneer, and the user is the bidder, thereby increasing the flexibility of pricing.

[0150] Example 4

[0151] See Figure 6 This diagram illustrates the structure of a scheduling center according to Embodiment 4 of the present invention. It may include the following modules:

[0152] The task receiving module 401 is used to receive the model fine-tuning task sent by the user terminal. The model fine-tuning task includes the user's bid, the fine-tuning dataset, and the computational requirements.

[0153] The optimal scheduling combination determination module 402 is used to determine the optimal scheduling combination of the model fine-tuning task based on the bid, the fine-tuning dataset, and the computing requirements, combined with the node capability information of each computing node in the computing cluster, with the goal of maximizing total revenue. The optimal scheduling combination includes target computing nodes selected from the computing cluster that meet the computing requirements; the node capability information includes computing power, video memory capacity, and video memory bandwidth capacity; and the total revenue includes the user's revenue and the revenue of the intelligent computing service provider.

[0154] The task acceptance judgment module 403 is used to determine whether to accept the model fine-tuning task based on the optimal scheduling combination.

[0155] The task execution module 404 is used to control the target computing node to execute the model fine-tuning task based on the fine-tuning dataset if the model fine-tuning task is accepted.

[0156] In one embodiment of the present invention, the scheduling center further includes the following modules:

[0157] The actual cost determination module is used to calculate the actual payment cost based on the optimal scheduling combination;

[0158] The actual payment return module is used to return the actual payment amount to the user terminal so that the user can make payment according to the actual payment amount.

[0159] In one embodiment of the present invention, the optimal scheduling combination determination module 402 further includes the following modules:

[0160] The primal optimization problem modeling module, based on the bid, fine-tuning dataset, and computational requirements, and combined with the node capability information of each computing node in the computing cluster, models a primal optimization problem based on decision variables with the objective of maximizing total revenue. The decision variables include variables related to data providers, computing nodes, and task acceptance. The data provider is a preprocessing service provider that preprocesses the task before the computing nodes process it. The primal optimization problem includes a first set of constraints.

[0161] An equivalence problem modeling module is used to simplify the original optimization problem into an equivalence problem based on scheduling combination, wherein the scheduling combination is a combination of execution elements of the decision variables that satisfy the constraints of the first set of constraints on the task itself, and the execution elements include at least computing node information, data provider information, and task acceptance information.

[0162] The Lagrange dual problem generation module is used to generate the Lagrange dual problem corresponding to the original optimization problem based on the original dual algorithm and the equivalent problem.

[0163] The optimal combination determination module is used to determine the optimal scheduling combination of the model fine-tuning tasks based on the Lagrange dual problem.

[0164] In one embodiment of the present invention, the Lagrange dual problem includes multiple dual variables, which include variables related to the computational overhead, memory overhead, and memory bandwidth overhead of the computing node when performing the task; then the optimal combination determination module is specifically used for:

[0165] Based on the Lagrange dual problem, determine the values ​​of the dual variables corresponding to each scheduling combination;

[0166] Based on the values ​​of the dual variables corresponding to each scheduling combination, determine the functional representation of each scheduling combination;

[0167] The optimal scheduling combination is determined by using the maximum value independent variable point set function based on the functional representation of each scheduling combination.

[0168] In one embodiment of the present invention, the actual cost determination module is specifically used for:

[0169] The actual payment cost is calculated based on the values ​​of the dual variables corresponding to the optimal scheduling combination and the preset pricing function.

[0170] In one embodiment of the present invention, the total revenue is represented as follows:

[0171]

[0172] in,

[0173]

[0174] Where U represents the total revenue, U r U represents the revenue earned by all users. c U represents the revenue earned by the intelligent computing service provider. i b represents the benefit gained by the user who initiates model fine-tuning task i. i p represents the bid of the user who initiated the model fine-tuning task i. i u represents the actual cost paid by the user after model fine-tuning task i is accepted. i z represents the task acceptance information indicating whether model fine-tuning task i is accepted. in This indicates whether model fine-tuning task i requires data provider n for data preprocessing information, q. in x represents the fee charged by the data provider. ikt Indicates whether model fine-tuning task i,e is executed on computing node k during time period t. ikt This represents the system overhead caused by computing node k training model fine-tuning task i within time period t.

[0175] In one embodiment of the present invention, the original optimization problem is expressed as:

[0176]

[0177]

[0178] Where P represents maximizing total revenue, and constraints (a1)-(a9) form the first set of constraints;

[0179] (a1) is used to ensure that for each model fine-tuning task i, if the task is accepted and requires data preprocessing, at most one data provider is selected; f i Indicates whether model fine-tuning task i requires data preprocessing;

[0180] (a2) is used to ensure that each model fine-tuning task i runs on at most one computing node k in each time period t;

[0181] (a3) is used to ensure that each model fine-tuning task i is executed only after it arrives at the system and completes data preprocessing; a i h represents the arrival time of model fine-tuning task i; in This represents the latency required by the data provider to process the training dataset for model fine-tuning task i;

[0182] (a4) is used to ensure that each model fine-tuning task i is executed before its deadline; d i This indicates the deadline for model fine-tuning task i;

[0183] (a5) is used to ensure that each model fine-tuning task i has completed sufficient computation; M i This represents the computational requirement for model fine-tuning task i; s ik This represents the amount of computation that can be completed in each time interval when model fine-tuning task i is executed on computing node k;

[0184] (a6) is used to constrain the computational capacity of each time period and each computing node; C kp This represents the computational power of a node k∈[K].

[0185] (a7) is used to constrain the memory capacity for each time period and each computing node; C km This represents the video memory capacity of computing node k;

[0186] (a8) is used to constrain the memory bandwidth capacity for each time period and each computing node; C kg This represents the memory bandwidth capability of computing node k;

[0187] (a9) specifies the range of values ​​for each decision variable.

[0188] In one embodiment of the present invention, the equivalence problem is expressed as:

[0189]

[0190]

[0191]

[0192] Where P1 represents maximizing the total revenue after equivalence, and constraints (b1)-(b5) form the second set of constraints; ζ i This refers to the set of all scheduling combinations for model fine-tuning task i that satisfy the necessary constraints; the necessary constraints include constraints (a1) to (a5); b il x represents the increment of the objective function value of problem P when using scheduling combination l to execute model fine-tuning task i; il Indicates whether model fine-tuning task i is executed according to scheduling combination l; skt (il), r kt (il), σ kt (il) represent the computational cost, memory cost, and memory bandwidth cost when the model fine-tuning task i is executed on the computing node k using the scheduling combination l during the time period t.

[0193] In one embodiment of the present invention, the Lagrange duality problem is expressed as:

[0194]

[0195] Where, μ i , λ kt , γ kt Let (b1), (b2), (b3), and (b4) represent the dual variables associated with constraints (b1), (b2), (b3), and (b4), respectively.

[0196] The function is represented as follows:

[0197]

[0198] Where F(il) represents the functional representation of scheduling combination l; (k,t)∈l refers to x in scheduling combination l ikt For a computation node k with a time interval t equal to 1, there are one or more (k,t) pairs in a scheduling combination l.

[0199] The optimal scheduling combination is expressed as:

[0200]

[0201] Among them, l i This represents the optimal scheduling combination;

[0202] The step of determining whether to accept the model fine-tuning task based on the optimal scheduling combination includes:

[0203] μ i =max{0,F(il)}

[0204] If μ i If the value is 0, the model fine-tuning task is deemed unacceptable.

[0205] If μ i If F(il) > 0, then the model fine-tuning task is accepted;

[0206] The actual payment amount is calculated using the following formula:

[0207]

[0208] The scheduling center provided in this embodiment of the invention can execute the model fine-tuning method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of executing the model fine-tuning method.

[0209] Example 5

[0210] See Figure 7 This diagram illustrates a structural schematic of an electronic device according to an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, blade servers, mainframe computers, and other suitable computers. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0211] like Figure 7 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0212] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0213] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as model fine-tuning methods based on intelligent computing platforms.

[0214] In some embodiments, the model fine-tuning method based on the intelligent computing platform can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the model fine-tuning method based on the intelligent computing platform described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the model fine-tuning method based on the intelligent computing platform by any other suitable means (e.g., by means of firmware).

[0215] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0216] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0217] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0218] Example 6

[0219] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the model fine-tuning method based on an intelligent computing platform as provided in any embodiment of this invention.

[0220] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0221] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A model fine-tuning method based on an intelligent computing platform, characterized in that, The intelligent computing platform includes a user terminal, a scheduling center, and a computing cluster. The method is applied to the scheduling center side and includes: Receive the model fine-tuning task sent by the user terminal, wherein the model fine-tuning task includes at least the user's bid, the fine-tuning dataset, and the computational requirements; Based on the bid, the fine-tuning dataset, and the computational requirements, and combined with the node capability information of each computing node in the computing cluster, the optimal scheduling combination for the model fine-tuning task is determined with the goal of maximizing total revenue. The optimal scheduling combination includes target computing nodes selected from the computing cluster that meet the computational requirements. The node capability information includes computing power, video memory capacity, and video memory bandwidth capacity. The total revenue includes revenue for the user and revenue for the intelligent computing service provider. Determine whether to accept the model fine-tuning task based on the optimal scheduling combination; If the model fine-tuning task is accepted, the target computing node is controlled to execute the model fine-tuning task based on the fine-tuning dataset; The step of determining the optimal scheduling combination of the model fine-tuning tasks based on the bid, the fine-tuning dataset, and the computational requirements, combined with the node capability information of each computing node in the computing cluster, with the goal of maximizing total revenue, includes: Based on the bid, the fine-tuning dataset, and the computational requirements, combined with the node capability information of each computing node in the computing cluster, a primal optimization problem based on decision variables is modeled with the goal of maximizing total revenue. The decision variables include variables related to the data provider, computing nodes, and task acceptance. The data provider is a preprocessing service provider that preprocesses the task before the computing nodes process it. The primal optimization problem includes a first set of constraints. The original optimization problem is simplified into an equivalent problem based on scheduling combination, wherein the scheduling combination is a combination of the decision variables and the execution elements that satisfy the constraints of the first set of constraints on the task itself. The execution elements include at least computing node information, data provider information and task acceptance information. Based on the primal dual algorithm and the equivalent problem, the Lagrange dual problem corresponding to the primal optimization problem is generated. The optimal scheduling combination of the model fine-tuning tasks is determined based on the Lagrange duality problem. The Lagrange dual problem includes multiple dual variables, which include variables related to the computational overhead, memory overhead, and memory bandwidth overhead of the computing node when performing the task. The determination of the optimal scheduling combination of the model fine-tuning tasks based on the Lagrange dual problem includes: Based on the Lagrange dual problem, determine the values ​​of the dual variables corresponding to each scheduling combination; Based on the values ​​of the dual variables corresponding to each scheduling combination, determine the functional representation of each scheduling combination; The optimal scheduling combination is determined by using the maximum value independent variable point set function based on the functional representation of each scheduling combination.

2. The method according to claim 1, characterized in that, After determining the optimal scheduling combination of the model fine-tuning tasks, the method further includes: Calculate the actual payment cost based on the optimal scheduling combination; The actual payment amount is returned to the user so that the user can make payment according to the actual payment amount.

3. The method according to claim 1, characterized in that, The step of calculating the actual payment cost based on the optimal scheduling combination includes: The actual payment cost is calculated based on the values ​​of the dual variables corresponding to the optimal scheduling combination and the preset pricing function.

4. The method according to claim 1 or 3, characterized in that, The total revenue is represented as follows: in, in, Represents total revenue. This represents the revenue received by all users. This indicates the revenue earned by the intelligent computing service provider. Indicates initiating a model fine-tuning task i The benefits received by users Indicates initiating a model fine-tuning task i The user's bid, This indicates the model fine-tuning task. i The actual payment made by the user after acceptance This indicates the model fine-tuning task. i Whether the task has been accepted or not, information received. This indicates the model fine-tuning task. i Does data provider n require data preprocessing information? This indicates the fees charged by the data provider. Indicates whether the model fine-tuning task is executed on computing node k during time period t. i , This indicates that computing node k is for the model fine-tuning task. i The system overhead caused by training within time period t.

5. The method according to claim 4, characterized in that, The primal optimization problem based on decision variables, modeled with the goal of maximizing total revenue based on the bid, is expressed as: in, This represents maximizing total revenue, with the following constraints. Form the first set of constraints; This is used to ensure that for each model fine-tuning task i, if the task is accepted and requires data preprocessing, at most one data provider can be selected. Indicates whether model fine-tuning task i requires data preprocessing; This is to ensure that each model fine-tuning task i runs on at most one compute node k in each time period t; This is to ensure that each model fine-tuning task i is executed only after it arrives at the system and has completed data preprocessing; This indicates the arrival time of model fine-tuning task i; This represents the latency required by the data provider to process the training dataset for model fine-tuning task i; This is used to ensure that each model fine-tuning task i is executed before its deadline; This indicates the deadline for model fine-tuning task i; This is used to ensure that each model fine-tuning task i completes a sufficient amount of computation; This represents the computational requirements of model fine-tuning task i; This represents the amount of computation that can be completed in each time interval when model fine-tuning task i is executed on computing node k; Used to constrain the computing capacity of each time period and each computing node; This represents the computational power of a node k∈[K]. Used to constrain the memory capacity of each time period and each computing node; This represents the video memory capacity of computing node k; Used to constrain the memory bandwidth capacity of each time period and each computing node; This represents the memory bandwidth capability of computing node k; The range of values ​​for each decision variable is specified.

6. The method according to claim 5, characterized in that, The equivalence problem is expressed as: in, This represents maximizing the total revenue after the equivalent process, with the following constraints. Form the second set of constraints; It refers to the set of all scheduling combinations for model fine-tuning task i that satisfy the necessary constraints; the necessary constraints include constraints ; Indicates the use of scheduling combination The increment of the objective function value of problem P when performing model fine-tuning task i; Indicates whether model fine-tuning task i follows the scheduling combination. implement; , , These represent the scheduling combinations used on computing node k during time period t. The computational cost, memory cost, and memory bandwidth cost when performing model fine-tuning task i.

7. The method according to claim 6, characterized in that, The Lagrange duality problem is expressed as: in, , Representing and constraint respectively Related dual variables; The functional representation of each scheduling combination, determined based on the values ​​of the dual variables corresponding to each scheduling combination, is expressed as follows: in, Represents scheduling combination The function representation; Refers to scheduling combination middle The computation node k and time period t Yes, in a scheduling combination There are one or more right; The optimal scheduling combination is determined by using the maximum value independent variable point set function based on the functional representation of each scheduling combination, as follows: in, This represents the optimal scheduling combination; The step of determining whether to accept the model fine-tuning task based on the optimal scheduling combination includes: like If so, the model fine-tuning task is deemed unacceptable. like If so, it is determined that the model fine-tuning task is accepted; The actual payment cost is calculated based on the optimal scheduling combination using the following formula: 。 8. An intelligent computing platform, characterized in that, The intelligent computing platform includes a user terminal, a scheduling center, and a computing cluster. The scheduling center includes the following modules: The task receiving module is used to receive the model fine-tuning task sent by the user terminal. The model fine-tuning task includes the user's bid, the fine-tuning dataset, and the computational requirements. The optimal scheduling combination determination module is used to determine the optimal scheduling combination for the model fine-tuning task based on the bid, the fine-tuning dataset, and the computing requirements, combined with the node capability information of each computing node in the computing cluster, with the objective of maximizing total revenue. The optimal scheduling combination includes target computing nodes selected from the computing cluster that meet the computing requirements; the node capability information includes computing power, video memory capacity, and video memory bandwidth capacity; and the total revenue includes the revenue of the user and the revenue of the intelligent computing service provider. The task acceptance judgment module is used to determine whether to accept the model fine-tuning task based on the optimal scheduling combination. The task execution module is used to control the target computing node to execute the model fine-tuning task based on the fine-tuning dataset if the model fine-tuning task is accepted. The optimal scheduling combination determination module includes: The primal optimization problem modeling module, based on the bid, fine-tuning dataset, and computational requirements, and combined with the node capability information of each computing node in the computing cluster, models a primal optimization problem based on decision variables with the objective of maximizing total revenue. The decision variables include variables related to data providers, computing nodes, and task acceptance. The data provider is a preprocessing service provider that preprocesses the task before the computing nodes process it. The primal optimization problem includes a first set of constraints. An equivalence problem modeling module is used to simplify the original optimization problem into an equivalence problem based on scheduling combination, wherein the scheduling combination is a combination of execution elements of the decision variables that satisfy the constraints of the first set of constraints on the task itself, and the execution elements include at least computing node information, data provider information, and task acceptance information. The Lagrange dual problem generation module is used to generate the Lagrange dual problem corresponding to the original optimization problem based on the original dual algorithm and the equivalent problem. The optimal combination determination module is used to determine the optimal scheduling combination of the model fine-tuning tasks based on the Lagrange dual problem. The Lagrange dual problem includes multiple dual variables, which include variables related to the computational overhead, memory overhead, and memory bandwidth overhead of the computing node when performing the task; therefore, the optimal combination determination module is specifically used for: Based on the Lagrange dual problem, determine the values ​​of the dual variables corresponding to each scheduling combination; Based on the values ​​of the dual variables corresponding to each scheduling combination, determine the functional representation of each scheduling combination; The optimal scheduling combination is determined by using the maximum value independent variable point set function based on the functional representation of each scheduling combination.

Images