An intelligent scheduling method and system based on an artificial intelligence large model

Abstract

The application provides an intelligent scheduling method and system based on an artificial intelligence large model, the method takes a server as an execution subject, and the method comprises the following steps: S1: collecting hardware feature data of a heterogeneous hardware cluster and task / instruction feature data of a large model task / instruction to be scheduled, and storing the hardware feature data and the task / instruction feature data into a local database. The intelligent scheduling method and system based on the artificial intelligence large model take the server as the only execution subject, follow a computer data processing core logic of data collection-data processing-model inference-decision generation-real-time monitoring-dynamic adjustment-fault tolerance rescheduling, break through technical barriers of existing large model scheduling technologies in terms of heterogeneous hardware adaptation and predictive scheduling by constructing a hardware-task dynamic mapping model and a timing prediction scheduling model, and realize full-process automation, intelligentization and refinement of large model scheduling.

Description

Technical Field

[0001] This invention relates to the field of big data service technology, and in particular to an intelligent scheduling method and system based on an artificial intelligence big data model. Background Technology

[0002] With the rapid development of artificial intelligence technology, the parameter scale and computational demands of large models are growing exponentially, placing extremely high demands on the scheduling efficiency of computing resources and hardware adaptability. The training and inference tasks of large models require the collaborative support of various heterogeneous hardware such as CPUs, GPUs, TPUs, and NPUs, while also facing challenges such as sudden task bursts, variable resource requirements, and complex instruction flows. Therefore, efficient intelligent scheduling methods have become one of the core key technologies for the practical application of large models.

[0003] There are several existing scheduling optimization methods for large models. Among them, prior art document 1, announcement number CN119311395B, discloses "a method for optimizing the scheduling of computing resources for large models", which divides tasks into abnormal tasks and normal tasks, and computing nodes into normal computing nodes and buffered computing nodes. It allocates resources based on node weight and connection number, and prioritizes the processing of abnormal tasks. Prior art document 2, announcement number CN119473560B, discloses "an instruction optimization scheduling method based on large models", which converts the model into a computing task graph, determines the scheduling priority based on the importance and timeliness of instruction paths, and dynamically adjusts the scheduling strategy in combination with register pressure status.

[0004] However, the aforementioned existing technologies still have significant technical shortcomings and cannot meet the core requirements of large-scale artificial intelligence models for intelligent scheduling, specifically in the following aspects: First, Comparison Document 1 only abstracts the hardware into ordinary and buffered computing nodes, setting node weights based on resource processing capabilities and request counts, without deeply binding them to the underlying hardware characteristics such as registers, computing units, and storage bandwidth. Hardware adaptation remains at a coarse-grained mapping at the node level. While Comparison Document 2 incorporates VLIW processors, registers, and dedicated computing units, hardware adaptation is limited to specific homogeneous computing units, lacking a general adaptation mechanism for heterogeneous hardware, and failing to consider the interaction of hardware resources between computing, storage, and network. Neither document establishes a dynamic mapping model between hardware characteristics and scheduling strategies, resulting in virtually zero adaptability to hardware failures, computing power fluctuations, and mixed deployments of heterogeneous hardware. It cannot achieve dynamic collaborative adaptation across the entire hardware stack, leading to insufficient deep collaborative adaptation capabilities for existing hardware resources and weak versatility for heterogeneous hardware.

[0005] Secondly, in contrast, document 1 handles sudden abnormal tasks by passively converting idle ordinary nodes into buffered temporary nodes, without predicting the probability of abnormal tasks, resource requirements, or duration. This reactive approach can easily lead to delayed resource allocation. Document 2 manages register pressure only through real-time monitoring and assessment, failing to anticipate sudden changes in instruction flow (such as sudden changes in instruction complexity, sudden data dependencies, or surges in register demand). The scheduling strategy is only adjusted after the pressure exceeds a threshold, easily leading to register resource conflicts. Neither document has an active scheduling module that combines time-series prediction and big data mining. Their response to the unknown and sudden demands in large-scale model scheduling is limited to post-event remediation, failing to achieve advance prediction and proactive resource allocation. This results in a lack of predictive scheduling capabilities in existing technologies, making their response to unknown demands overly passive. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the existing technology. The present invention proposes an intelligent scheduling method and system based on a large artificial intelligence model.

[0007] To address the aforementioned issues of lacking a dynamic mapping model between hardware characteristics and scheduling strategies, resulting in near-zero adaptability to hardware failures, computing power fluctuations, and heterogeneous hardware deployments, and failing to achieve dynamic collaborative adaptation across the entire hardware stack, the current technology suffers from insufficient deep collaborative adaptation capabilities for hardware resources and weak versatility for heterogeneous hardware. Furthermore, the absence of an active scheduling module combining time-series prediction and big data mining means that responses to unknown and sudden demands in large-scale model scheduling are limited to post-event remediation, failing to achieve advance prediction and proactive resource allocation. This leads to a lack of predictive scheduling capabilities in existing technologies and a passive approach to handling unknown demands. Therefore, the technical solution adopted in this invention is: A method and system for intelligent scheduling based on a large-scale artificial intelligence model, wherein the method uses a server as the execution entity, and the method includes: S1: Collect hardware feature data of heterogeneous hardware clusters and task / instruction feature data of large model tasks / instructions to be scheduled, and store the hardware feature data and task / instruction feature data in a local database; S2: Retrieve the hardware feature data and task / instruction feature data from the local database, and perform cleaning, normalization and feature quantization processing on the two types of data respectively to obtain standardized hardware feature vectors and standardized task / instruction feature vectors; S3: Construct a hardware-task dynamic mapping model, input the standardized hardware feature vector and the standardized task / instruction feature vector into the hardware-task dynamic mapping model, and calculate the matching degree value between hardware and task / instruction; Construct a timing prediction scheduling model, input the preprocessed timing monitoring data into the timing prediction scheduling model, and output the resource requirement prediction value, hardware state change prediction value, and probability value of sudden task occurrence for task / instruction. S4: Based on the matching degree value, resource demand prediction value, hardware status change prediction value, and probability value of sudden task occurrence, combined with preset scheduling constraints, an initial intelligent scheduling decision is generated. The initial intelligent scheduling decision includes hardware allocation results, task / instruction execution priority, and resource configuration parameters. S5: Real-time acquisition of hardware operation data and task / instruction execution data during the scheduling and execution process; feature extraction of the data to obtain a real-time monitoring feature vector; calculation of the difference between the real-time monitoring feature vector and the predicted values ​​of hardware state changes and resource requirements to obtain a deviation value. S6: Determine whether the deviation value exceeds the preset deviation threshold. If it does not exceed the threshold, maintain the initial intelligent scheduling decision. If it exceeds the threshold, input the real-time monitoring feature vector into the time-series prediction scheduling model for online update to obtain the updated prediction value. Based on the updated prediction value and the matching degree value, re-optimize the scheduling decision and send it to the heterogeneous hardware cluster for execution. S7: Real-time detection of hardware fault data and abnormal execution data of tasks / instructions in heterogeneous hardware clusters to determine whether there is a hardware fault or abnormal execution of tasks / instructions. If not, the current scheduling decision continues to be executed; if so, the matching degree is recalculated based on the hardware-task dynamic mapping model to generate a fault-tolerant rescheduling decision, replace the faulty hardware allocation result and reallocate the abnormal tasks / instructions.

[0008] Preferably, in step S1, the hardware feature data includes hardware type dimension data, hardware performance dimension data, and hardware status dimension data. The hardware type dimension data includes the hardware identifier, computing unit type, and register configuration parameters of the CPU / GPU / TPU / NPU. The hardware performance dimension data includes floating-point operation capability, storage bandwidth, network transmission rate, and number of registers. The hardware status dimension data includes CPU utilization, GPU memory occupancy rate, register utilization rate, hardware temperature, and network latency. The task / instruction feature data includes task type dimension data, task attribute dimension data, and instruction feature dimension data. The task type dimension data includes task identifiers for large model training tasks, inference tasks, and data preprocessing tasks. The task attribute dimension data includes computational complexity, data size, task urgency, task importance, and execution time requirements. The instruction feature dimension data includes the number of instruction registers required, data dependency depth, computational unit affinity, and instruction execution time.

[0009] Preferably, in step S2, feature quantization processing is performed on the hardware feature data and task / instruction feature data, specifically as follows: The hardware feature data is quantized using a linear weighting method to obtain the quantized hardware feature values. The calculation formula is as follows: in, For the first The characteristic quantization value of each hardware component The number of dimensions of hardware features. For the first The weight coefficients of each hardware feature dimension, and =1, For the first The first hardware Normalized values ​​for each feature dimension; The task / instruction feature data is quantized using a linear weighting method to obtain the quantized value of the task / instruction feature. The calculation formula is as follows: in, For the first The characteristic quantization value of each task / instruction The number of task / instruction feature dimensions. For the first The weight coefficients of each task / instruction feature dimension, and =1, For the first The first task / instruction Normalized values ​​for each feature dimension; The quantized values ​​of the hardware features and the quantized values ​​of the task / instruction features are subjected to vector normalization to obtain standardized hardware feature vectors and standardized task / instruction feature vectors. The normalization formula is as follows: , in, For the first The standardized hardware feature vector of a hardware unit. For the first A standardized task / instruction feature vector for each task / instruction. , These are the L2 norms of the original hardware feature vector and the original task / instruction feature vector, respectively.

[0010] Preferably, in step S3, constructing a hardware-task dynamic mapping model and calculating the matching degree value between hardware and task / instruction specifically involves: A matching degree calculation module for the hardware-task dynamic mapping model is constructed using the cosine similarity algorithm. The calculation formula is as follows: in, For the first The hardware and the first The matching degree value of each task / instruction is set, with a value ranging from [0,1]. A larger value indicates a higher matching degree. The matching degree value is then corrected based on a collaborative filtering algorithm to obtain the corrected matching degree value. The calculation formula is: in, This is a correction factor, with a value range of [0, 0.5]. In order to be with the first The number of hardware devices of the same type For the first The hardware and the first Similarity values ​​for similar hardware.

[0011] Preferably, in step S3, constructing the time-series predictive scheduling model specifically involves: A time-series predictive scheduling model is constructed using an LSTM neural network combined with an attention mechanism. The input to the model is the feature matrix of the time-series monitoring data. ,in for The monitoring feature vector at any given time; The hidden layer output of the LSTM is weighted using an attention mechanism. The formula for calculating the attention weights is as follows: in, for Time Hidden Layer Output Attention weights For trainable weight matrix, This is the bias term; the model's fusion output is: fused output The input is fed into the fully connected layer to obtain the prediction result. The loss function of the model is the mean squared error loss function, and the calculation formula is as follows: Where N is the number of training samples, For the true value, These are the model's predicted values.

[0012] Preferably, in step S4, generating the initial intelligent scheduling decision specifically involves: Based on the probability value P of the occurrence of the unexpected task, determine whether there is a risk of an unexpected task. ( To preset a sudden risk threshold, a preset proportion of hardware resources are reserved as buffer resources. The calculation formula is: Based on the corrected matching degree value Perform an initial matching of hardware with tasks / instructions, and allocate hardware in descending order of matching score; Based on the predicted resource requirements, the resources are configured for the matched task / instruction, including the number of register resources configured. The calculation formula is: in For the first Number of basic registers required per task / instruction For the predicted value of register resource requirements, This is a resource reservation coefficient, with a value range of [0, 0.2]. Quantification value based on task / instruction characteristics Calculate execution priority The calculation formula is: in, This is the priority weight coefficient, and , For the first The maximum corrected match value for a task / instruction.

[0013] Preferably, in step S5, calculating the deviation value specifically involves: calculating the hardware state deviation value respectively. Deviation value of resource demand The calculation formula is: in, For the number of hardware, For the first The actual state value of each piece of hardware. For the first The state prediction value of each hardware component. For the first The nominal status value of each hardware component; For the number of tasks / instructions, For the first The actual resource requirements of each task / instruction. For the first Forecasted resource requirements for each task / instruction. For the first The nominal value of resource requirements for each task / instruction; Calculate the overall deviation value The calculation formula is: in, This is the weighting coefficient for hardware status deviation, with a value range of [0.4, 0.6].

[0014] Preferably, in step S7, real-time detection of hardware fault data and abnormal execution data of tasks / instructions specifically involves: using a threshold detection method combined with trend analysis to detect hardware faults, and determining the hardware fault judgment value. The calculation formula is: in, For hardware fault monitoring indicators, For the first The first hardware The actual value of each monitoring indicator For the first Normal thresholds for each monitoring indicator This is the trend coefficient. The rate of change of hardware status indicators; like ( (If the preset fault determination threshold is used), then the fault determination threshold is determined. A hardware component malfunctioned; Execution schedule deviation detection is used to identify task / instruction execution anomalies, and the anomaly judgment value is... The calculation formula is: in, For the first The actual execution progress of each task / instruction. This refers to the actual execution time. To preset the execution schedule, The preset execution time; like ( (If the preset anomaly detection threshold is used), then the first anomaly is determined. An error occurred during the execution of a task / instruction.

[0015] Preferably, in step S7, generating a fault-tolerant rescheduling decision specifically involves: if a hardware failure is detected, removing the faulty hardware and selecting a corrected matching degree value from the hardware cluster. The second-highest priority hardware that is idle is used as a replacement hardware to reallocate unfinished tasks / instructions on the faulty hardware; if an abnormal task / instruction execution is detected, the resource requirement deviation value of the abnormal task / instruction is calculated. If the deviation value is caused by insufficient resources, the resource allocation is increased based on the updated resource requirement prediction value; if the deviation value is caused by low hardware matching degree, the hardware-task matching process is re-executed to select hardware with a higher matching degree for reallocation.

[0016] An intelligent scheduling system based on a large-scale artificial intelligence model, the system being deployed on a server, comprising: The data acquisition module is used to collect hardware feature data of heterogeneous hardware clusters and task / instruction feature data of large model tasks / instructions to be scheduled, and store the hardware feature data and task / instruction feature data in a local database. The data preprocessing module is used to retrieve the hardware feature data and task / instruction feature data from the local database, and perform cleaning, normalization and feature quantization on the two types of data to obtain standardized hardware feature vectors and standardized task / instruction feature vectors. The model building and inference module is used to build a hardware-task dynamic mapping model, calculate the matching degree between hardware and tasks / instructions, and build a timing prediction scheduling model to output the predicted resource requirements, hardware state changes, and probability of sudden tasks for tasks / instructions. The scheduling decision generation module is used to generate an initial intelligent scheduling decision based on the matching degree value, the predicted value of resource demand, the predicted value of hardware status change, and the probability value of sudden task occurrence, combined with preset scheduling constraints. The real-time monitoring module is used to collect hardware operation data and task / instruction execution data in real time during the scheduling process, extract real-time monitoring feature vectors, calculate deviation values, and determine whether they exceed the preset deviation threshold. The dynamic adjustment module is used to update the time-series prediction scheduling model online if the deviation value exceeds the preset deviation threshold, and then re-optimize the scheduling decision based on the updated prediction value and matching degree value before issuing it for execution. The fault-tolerant rescheduling module is used to detect hardware fault data and abnormal execution data of tasks / instructions in real time. If a fault or abnormality is found, it generates a fault-tolerant rescheduling decision and executes it.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention establishes a multi-dimensional hardware and task / instruction feature quantification system, realizing the digital expression of underlying hardware characteristics and task requirements; designs a matching degree calculation method combining cosine similarity and collaborative filtering, achieving refined matching between heterogeneous hardware and tasks / instructions; employs an LSTM+attention mechanism to construct a time-series prediction model, enabling advance prediction of task requirements and hardware status, upgrading from passive scheduling to active scheduling; establishes a real-time monitoring and dynamic adjustment mechanism based on comprehensive deviation values, allowing scheduling strategies to adapt to dynamic changes in hardware and tasks in real time; and designs a quantified hardware fault and task anomaly detection method, combined with the matching model, to achieve rapid fault-tolerant rescheduling, ensuring scheduling stability.

[0018] The method and system of this invention can be widely applied to scheduling scenarios for various artificial intelligence tasks, such as large-scale model training, inference, and data preprocessing. It supports the hybrid deployment of various heterogeneous hardware such as CPUs, GPUs, TPUs, and NPUs without requiring modifications to the underlying hardware, resulting in low deployment costs, strong compatibility, and good scalability. Practical applications show that this invention can improve heterogeneous hardware resource utilization by more than 30%, reduce task response latency by more than 40%, increase scheduling success rate to over 99%, and reduce register resource conflict rate to below 1.5%. It provides core technical support for the large-scale, engineering-oriented application of large models and has significant theoretical and practical value. Attached Figure Description

[0019] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein: Figure 1 This is the flowchart of the intelligent scheduling method of the present invention; Figure 2 The diagram below shows the construction block of the intelligent scheduling system of the present invention. Detailed Implementation

[0020] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0021] Specific embodiments of the present invention are described below with reference to the accompanying drawings.

[0022] Please see Figures 1-2 This embodiment proposes an intelligent scheduling method based on a large artificial intelligence model. The method is executed by a server, which is a computer device with data processing, model training, and network communication capabilities. It is equipped with a CPU, large-capacity memory, solid-state drive, and deployed with deep learning frameworks (such as TensorFlow and PyTorch) and database management systems (such as MySQL and Redis). The server communicates with a heterogeneous hardware cluster composed of CPU, GPU, TPU, and NPU through high-speed Ethernet or InfiniBand network to realize real-time data interaction and the issuance of scheduling instructions.

[0023] Step S1: Feature Data Acquisition and Storage The server collects hardware characteristic data of the heterogeneous hardware cluster and task / instruction characteristic data of the large model tasks / instructions to be scheduled through the data acquisition module. The acquisition method is as follows: hardware operation data is obtained from the hardware layer through hardware monitoring tools (such as nvidia-smi, ipmitool), and the configuration and operation data of tasks / instructions are obtained through the large model scheduling platform. The acquisition frequency is 100ms / time to ensure the real-time performance of the data.

[0024] Hardware feature data acquisition: Hardware feature data is divided into hardware type dimension, hardware performance dimension, and hardware status dimension. The specific acquisition indicators are shown in Table 1 below: Table 1 Hardware Feature Data Acquisition Indicators Task / Instruction Feature Data Collection: Task / instruction feature data is divided into task type dimension, task attribute dimension, and instruction feature dimension. Specific collection indicators are shown in Table 2 below. Table 2 Task / Instruction Feature Data Collection Indicators The values ​​for task urgency and task importance are both in the range of [1,5], with larger values ​​indicating higher urgency / importance; computational unit affinity refers to the type of hardware computational unit that the instruction adapts to, such as matrix multiplication unit or activation function unit, and is represented by the value [0,1] to indicate the degree of adaptation, with larger values ​​indicating higher affinity.

[0025] Data storage: The server stores the collected hardware feature data and task / instruction feature data in the format of timestamp-hardware ID / task ID-feature index-value to the local database. The structured data is stored in a MySQL relational database, and the real-time time series data is stored in a Redis cache database for easy retrieval and processing later.

[0026] Step S2: Feature Data Preprocessing The server retrieves hardware feature data and task / instruction feature data from the local database through the data preprocessing module. It then performs data cleaning, normalization, and feature quantization on the two types of data in sequence to eliminate data noise and the influence of units, resulting in standardized hardware feature vectors and standardized task / instruction feature vectors.

[0027] Data cleaning: The following rules are used to clean the data: Remove hardware / task feature data with more than 30% missing values; Linear interpolation is used to supplement feature data with a missing value ratio of ≤30%. The formula is as follows: ( (The number of intervals between the first valid data after the missing value). Outliers are removed using the 3σ principle; if the eigenvalues... satisfy ( The characteristic mean, If the value is the characteristic standard deviation, it is considered an outlier and replaced with the mean.

[0028] Data normalization: The cleaned feature data undergoes min-max normalization to map the feature values ​​to the [0,1] interval, eliminating the influence of dimensions. The formula is: in, These are the original eigenvalues. These are the normalized eigenvalues. , These are the minimum and maximum values ​​for that feature dimension, respectively.

[0029] Feature quantization: Hardware feature quantization: The normalized hardware feature data is quantized using a linear weighting method to obtain the quantized hardware feature value. The formula is as follows: in, The number of hardware feature dimensions (in this embodiment) ), Let be the weight coefficient of the k-th hardware feature dimension, determined by the Analytic Hierarchy Process (AHP), and satisfy . =1; In this embodiment, the hardware performance dimension accounts for 60% of the weight, the hardware status dimension accounts for 30% of the weight, and the hardware type dimension accounts for 10% of the weight. For the first The first hardware Normalized values ​​for each feature dimension.

[0030] Task / Instruction Feature Quantization: The normalized task / instruction feature data is quantized using a linear weighting method to obtain the quantized task / instruction feature values. The formula is as follows: in, The number of task / instruction feature dimensions (in this embodiment) ), For the first The weight coefficients for each task / instruction feature dimension are determined by the Analytic Hierarchy Process (AHP) and satisfy the following conditions: =1; In this embodiment, the weight of the task attribute dimension accounts for 50%, the weight of the instruction feature dimension accounts for 40%, and the weight of the task type dimension accounts for 10%. For the first Normalization of the m-th feature dimension of a task / instruction.

[0031] Vector standardization: The quantized values ​​of hardware features and task / instruction features are used to construct the original feature vector. The original feature vector is then subjected to L2 normalization to obtain the standardized hardware feature vector. and standardized task / instruction feature vectors The formula is: , in, , These are the L2 norms of the original hardware feature vector and the original task / instruction feature vector, respectively; the L2 norm of the normalized feature vector is 1, which facilitates subsequent similarity calculation.

[0032] Step S3: Model Building and Inference The server constructs a hardware-task dynamic mapping model and a timing prediction scheduling model through the model building and inference module, completes the training and inference of the model, and outputs the matching degree value between hardware and task / instruction, the predicted value of resource requirements of task / instruction, the predicted value of hardware state change, and the probability value of sudden task occurrence.

[0033] Hardware-Task Dynamic Mapping Model Construction and Inference The hardware-task dynamic mapping model is used to calculate the matching degree between hardware and tasks / instructions, enabling fine-grained matching between heterogeneous hardware and tasks / instructions. The model consists of a basic matching degree calculation module and a matching degree correction module. Basic matching degree calculation: The basic matching degree between hardware and task / instruction is calculated using the cosine similarity algorithm. Since the L2 norm of the standardized feature vector is 1, the formula simplifies to: in, For the first The hardware and the first The basic matching degree value of each task / instruction ranges from [0,1], with a larger value indicating a higher matching degree. The first normalized hardware feature vector Each dimension value The first feature vector of the standardized task / instruction Each dimension value (the number of feature dimensions is uniformly set) ).

[0034] Matching Degree Correction: The basic matching degree value is corrected based on the collaborative filtering algorithm, taking into account the impact of the similarity of similar hardware on the matching degree. The formula is as follows: in, As a correction factor, the value in this embodiment is 0.3; In order to be with the first The number of hardware devices of the same type; For the first The hardware and the first The similarity value of the same type of hardware is calculated by the cosine similarity algorithm; the corrected matching value... The value range remains [0,1], further improving the accuracy of the matching.

[0035] Model inference: The standardized hardware feature vector and standardized task / instruction feature vector are input into the trained hardware-task dynamic mapping model, and the corrected matching degree matrix of all hardware and tasks / instructions is output. ( For the number of hardware, (Number of tasks / instructions), stored in the local database.

[0036] Construction and Inference of Time-Series Predictive Scheduling Model The time-series predictive scheduling model is used to predict the resource requirements of tasks / instructions, hardware state changes, and the probability of sudden task occurrences. The model is built using an LSTM neural network and an attention mechanism, which can accurately capture the long-term dependencies and key features of time-series monitoring data. The training and inference process of the model is as follows: Training dataset construction: Time-series monitoring data from the past 30 days was selected as the training dataset, including hardware status time-series data, task / instruction resource requirement time-series data, and sudden task record data. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. A sliding window was applied to the data with a window size of 60 and a stride of 10 to construct the input feature matrix. ( For time steps, (For feature dimensions).

[0037] Model structure construction: The model is built based on the PyTorch framework. The model structure includes: Input layer: receives the feature matrix Feature Dimension (This embodiment); LSTM layer: Set up 3 LSTM layers with hidden layer dimensions of 256, 128 and 64 respectively, and use a Dropout layer (dropout=0.2) to prevent overfitting; Attention layer: The hidden layer output of the last layer of the LSTM is weighted to obtain the attention weights. ; Fusion Layer: Summing the weighted hidden layer outputs yields the fused output. ; Fully connected layers: Two fully connected layers are set up, with an output layer dimension of 3, corresponding to the predicted value of resource demand, the predicted value of hardware state change, and the probability value of sudden task occurrence, respectively.

[0038] Model training: The optimizer used was AdamW, with a learning rate of 0.001 and a weight decay of 0.0001. The loss function used is the mean squared error loss function, and the formula is: in, The number of training samples, For the true value, These are the model's predicted values; The training run consists of 100 rounds, using an early stopping method. If the validation set loss does not decrease for 10 consecutive rounds, training is stopped and the optimal model parameters are saved.

[0039] Model inference: The preprocessed time series monitoring data is input into the trained time series prediction and scheduling model. Model output: Resource requirement forecasts for tasks / instructions This includes register requirements, storage requirements, and computing power requirements. Hardware state change prediction value This includes predicted values ​​for hardware utilization, temperature, and network latency within the next 5 seconds. Probability of unexpected tasks occurring The probability of an unexpected task occurring within the next 5 seconds, with a value ranging from [0,1]. The model inference time is ≤50ms, meeting the requirements of real-time scheduling.

[0040] Step S4: Initial Intelligent Scheduling Decision Generation The server uses the scheduling decision generation module to generate a decision based on the corrected matching degree value. Resource demand forecast Hardware status change prediction value Probability of unexpected tasks occurring Based on preset scheduling constraints (such as hardware load limit ≤ 80% and resource allocation fairness), an initial intelligent scheduling decision is generated, which includes three parts: hardware allocation result, task / instruction execution priority, and resource configuration parameters.

[0041] Emergency resource reservation: The probability value P of an emergency task is used to determine the emergency risk. In this embodiment, an emergency risk threshold is preset. =0.5: like Then calculate the resource reservation ratio. Reserved from heterogeneous hardware clusters A proportion of idle hardware resources are used as buffer resources to cope with sudden tasks; like If no buffer resources are reserved, all hardware resources will participate in normal scheduling.

[0042] Hardware allocation: based on the corrected matching degree matrix Hardware and tasks / instructions are matched and allocated according to the principle of matching degree from high to low and hardware load from low to high. The allocation rules are as follows: For each task / instruction Select the corrected matching degree value Maximum hardware load ≤ 80% As the optimal matching hardware; If the optimal matching hardware load is >80%, then select the hardware with the second highest matching degree, and so on, until hardware that meets the load requirements is found. After allocation is complete, update the hardware load status to avoid duplicate allocation.

[0043] Resource allocation: based on resource demand forecasts Configure resources for the matched task / instruction, with core resources configured as register resources, using the following formula: in, The base register requirements for a task / instruction are obtained by the instruction parsing tool; In this embodiment, the resource reservation coefficient is set to 0.15. At the same time, storage resources and network resources are configured for tasks / instructions. The number of storage resources configured is 1.2 times the data size, and the number of network resources configured meets the data transmission rate requirements.

[0044] Execution priority calculation: based on the quantification value of task / instruction characteristics. Maximum corrected matching degree value Probability of unexpected tasks occurring The execution priority is calculated using the following formula: in, In this embodiment, the priority weighting coefficient is used. , , ,and ; The value range is [0,2]. The larger the value, the higher the execution priority. The server allocates execution time slices to tasks / instructions in order of priority from high to low.

[0045] Decision delivery: The server encapsulates the initial intelligent scheduling decision into a scheduling instruction in the format of hardware ID-task ID-resource configuration-execution priority, and delivers it to each hardware node of the heterogeneous hardware cluster through a high-speed network. After receiving the instruction, the hardware node executes the scheduling of the task / instruction according to the decision.

[0046] Step S5: Real-time monitoring and deviation calculation The server collects hardware operation data and task / instruction execution data in real time during the scheduling process through a real-time monitoring module. The collection frequency is 50ms / time. The collected data is used to extract features to obtain real-time monitoring feature vectors, and hardware status deviation value, resource requirement deviation value, and comprehensive deviation value are calculated.

[0047] Real-time monitoring feature extraction: The collected hardware operation data and task / instruction execution data are cleaned and normalized in the same way as in step S2 to extract the real-time feature vectors of hardware status and resource requirements, which are then merged into a real-time monitoring feature vector. ( (This embodiment).

[0048] Deviation calculation: Hardware status deviation value: Calculate the average relative deviation between the actual and predicted status values ​​of each hardware component. The formula is: in, For the first The actual state values ​​of the hardware (such as utilization rate, temperature). The predicted values ​​of hardware state changes output by the model. For the first The nominal status values ​​of each hardware component (e.g., nominal hardware utilization is 100%, nominal temperature is 85℃). The value range is [0,1], and the larger the value, the greater the deviation of the hardware state.

[0049] Resource demand deviation: Calculate the average relative deviation between the actual and predicted resource demand values ​​for each task / instruction. The formula is: in, For the first The actual resource requirements of each task / instruction. The resource demand forecast output by the model is labeled as the [number]. The nominal value of resource requirements for each task / instruction; The value range is [0,1], and the larger the value, the greater the deviation in resource demand.

[0050] Overall Deviation Value: The overall deviation value is obtained by combining the hardware status deviation value and the resource requirement deviation value. The formula is as follows: in, This is the weighting coefficient for hardware status deviation. In this embodiment, it is set to 0.5 to ensure that the impact of deviations in hardware status and resource requirements is equal. The value range is [0,1].

[0051] Step S6: Dynamic Adjustment of Scheduling Decisions The server determines the overall deviation value through a dynamic adjustment module. Does it exceed the preset deviation threshold (the preset deviation threshold in this embodiment)? ), and perform the corresponding operation based on the judgment result: Maintain the initial scheduling decision: if This indicates that the deviation between the scheduling execution status and the prediction result is within the allowable range, and the server maintains the initial intelligent scheduling decision and continues to monitor the scheduling execution process in real time.

[0052] Update the model and optimize scheduling decisions: If This indicates that the scheduling execution status deviates significantly from the prediction result, requiring dynamic optimization of the scheduling decision. The steps are as follows: Online model updates: Real-time monitoring of feature vectors As new training samples, they are input into the time series prediction scheduling model for online fine-tuning. The model parameters are updated using the mini-batch gradient descent method (batch_size=16), and the fine-tuning time is ≤100ms. Re-forecasting: The updated online model is used to re-infer the current time-series monitoring data to obtain updated resource demand forecasts. Hardware status change prediction value ; Re-optimize scheduling decisions: based on updated predictions and corrected matching values. Following the scheduling decision generation rules in step S4, the optimized scheduling decision is regenerated, and the adjustments include hardware reallocation, resource configuration adjustment, and execution priority sorting update. Decision delivery: The server encapsulates the optimized scheduling decision into a scheduling instruction and sends it to the heterogeneous hardware cluster. After receiving the instruction, the hardware nodes adjust the scheduling execution strategy in real time to achieve dynamic optimization of the scheduling decision.

[0053] Step S7: Fault-tolerant rescheduling The server uses a fault-tolerant rescheduling module to detect hardware fault data and abnormal execution data of tasks / instructions in heterogeneous hardware clusters in real time. It uses fault judgment values ​​and abnormal judgment values ​​to accurately identify hardware faults and abnormal task / instruction execution, and generates fault-tolerant rescheduling decisions based on the identification results.

[0054] Hardware fault detection: A threshold detection method combined with trend analysis is used to calculate the hardware fault determination value. The formula is as follows: In this embodiment (Hardware fault monitoring metrics: utilization, temperature, network latency, register utilization, storage bandwidth); For the first The first hardware The actual value of each monitoring indicator For the first Normal thresholds for each monitoring indicator; This is the trend coefficient, and in this embodiment, it is set to 0.2. The rate of change of hardware status indicators is obtained through linear fitting. The range of values ​​is In this embodiment, a preset fault determination threshold is used. ,like Then determine the first A hardware component malfunctioned.

[0055] Task / Instruction Anomaly Detection: The task / instruction anomaly judgment value is calculated using execution progress deviation detection. The formula is as follows: in, For the first The actual execution progress (%) of a task / instruction is the actual execution time (s); The preset execution progress (%) and preset execution time (s) are given. The range of values ​​is In this embodiment, a preset anomaly detection threshold is provided. ,like Then determine the first An error occurred during the execution of a task / instruction.

[0056] Fault-tolerant rescheduling decision generation and execution: Hardware fault tolerance: If a hardware fault is detected, the server immediately removes the faulty hardware and selects a corrected matching value from the hardware cluster. Idle hardware with the second highest load and ≤70% load is used as replacement hardware; unfinished tasks / instructions on the faulty hardware are reassigned to the replacement hardware, resource configuration parameters are updated for the replacement hardware, and execution priority is increased by 10% to ensure rapid recovery and execution of tasks / instructions.

[0057] Task Fault Tolerance: If a task / instruction execution exception is detected, the server first calculates the resource requirement deviation of the exception task / instruction. : like (If resource shortage causes an anomaly), then the updated resource demand forecast will be used. Increase resource allocation by 20% for abnormal tasks / instructions while maintaining the original hardware allocation; like If the abnormality is caused by low hardware compatibility, then the hardware-task matching process in step S3 will be re-executed to select a corrected compatibility value for the abnormal task / instruction. Higher-level hardware is reallocated and execution priorities are adjusted.

[0058] Scheduling Recovery: After the fault-tolerant rescheduling decision is issued and executed, the server monitors the running status of the replacement hardware and the execution status of abnormal tasks / instructions in real time. When the hardware load recovers to the normal range (≤80%) and the task execution progress recovers to the preset range (the deviation between the actual execution progress and the preset execution progress is ≤10%), the scheduling is determined to be normal. The server stops the fault-tolerant rescheduling process, incorporates the replacement hardware into the regular hardware cluster management, restores the regular dynamic scheduling strategy for abnormal tasks / instructions, and continues to execute real-time monitoring and dynamic adjustment steps.

[0059] Fault and Anomaly Logs: The server stores relevant information about hardware faults (fault hardware ID, fault occurrence time, fault monitoring index value, fault recovery time) and relevant information about task / instruction execution anomalies (abnormal task ID, anomaly occurrence time, anomaly judgment value, anomaly recovery measures) in the local database to form fault / anomaly logs, providing data support for subsequent hardware maintenance, model parameter optimization, and scheduling strategy adjustment.

[0060] At this point, the server has completed a full intelligent scheduling process for a large model. The entire process, from feature data acquisition to fault-tolerant rescheduling, forms a closed loop. All steps are executed automatically by the server, and data exchange between modules is shared through a local database. The total time for a single round of scheduling is ≤500ms, meeting the real-time scheduling requirements for large model training and inference. Furthermore, the server supports continuous execution of multiple rounds of scheduling. When new tasks / instructions are added or hardware states change, the server will trigger a new round of scheduling, achieving full automation and intelligence in the scheduling of large models.

[0061] Example of use To verify the practical application effect of the intelligent scheduling method and system based on large artificial intelligence models of the present invention, this embodiment selects a large model inference platform of an artificial intelligence technology company as the application scenario. This platform mainly undertakes online inference tasks of large models with hundreds of billions of parameters, and faces problems such as sudden inference tasks, difficulty in coordinating heterogeneous hardware resources, and low resource utilization of traditional scheduling methods. The method and system of the present invention are used to intelligently schedule the platform, and a comparative test is carried out with two scheduling methods in the prior art (comparison file 1: the computing resource scheduling method of CN119311395B, and comparison file 2: the instruction optimization scheduling method of CN119473560B) to verify the technical advantages of the present invention.

[0062] Test environment configuration The heterogeneous hardware cluster of the test platform consists of one server (the main scheduling and execution entity) and ten hardware nodes. The hardware nodes include three types of heterogeneous hardware: CPU, GPU, and NPU. The specific configuration is shown in Table 3 below. The server and hardware nodes communicate via 100Gbps high-speed Ethernet with a network latency of ≤0.5ms.

[0063] Table 3 Heterogeneous Hardware Cluster Configuration Table Software environment Operating System: Both the server and hardware nodes use Ubuntu 20.04 LTS 64-bit operating system; Scheduling System: Deploy the intelligent scheduling system based on the artificial intelligence large model of this invention, which includes seven modules: data acquisition, preprocessing, model building and inference, scheduling decision generation, real-time monitoring, dynamic adjustment, and fault-tolerant rescheduling; Comparison scheduling system: Deploy the software systems corresponding to the scheduling methods in comparison file 1 and comparison file 2 respectively; Large Model and Tasks: A large Chinese language model (LLM) with hundreds of billions of parameters is used. The test task is an online reasoning task of the large model, which includes ordinary reasoning tasks (text generation, question answering interaction) and sudden reasoning tasks (batch text processing, multimodal reasoning). The specific characteristics of the tasks are shown in Table 4 below.

[0064] Table 4: Characteristics of Inference Tasks in Major Models Test metrics Four core metrics were selected as the test evaluation criteria: hardware resource utilization, task response latency, scheduling success rate, and register resource conflict rate. The definitions of each metric are as follows: Hardware resource utilization: The average computing resource utilization of CPU, GPU, and NPU in a heterogeneous hardware cluster, calculated using the following formula: ; Task response latency: The time interval from task submission to the start of execution on the hardware node, in milliseconds (ms), which is the average response latency of all tasks; Scheduling success rate: The percentage of tasks that are successfully executed and completed out of the total number of tasks. The formula is as follows: ; Register resource contention rate: The proportion of tasks that encounter register resource contention out of the total number of tasks. The calculation formula is as follows: .

[0065] Testing process This test was divided into three phases, using the scheduling method of this invention, the scheduling method of Comparative Document 1, and the scheduling method of Comparative Document 2 to schedule the large model inference task on the test platform. Each phase lasted for 2 hours. During the test, the hardware environment, software environment, and task characteristics were kept consistent, and only the scheduling method was adjusted. The specific test process is as follows: Testing process of the scheduling method of the present invention Data Acquisition and Preprocessing: The scheduling server collects hardware feature data of the heterogeneous hardware cluster and task / instruction feature data of the inference task at a frequency of 100ms / time through the data acquisition module. The data is cleaned, normalized and feature quantized to obtain standardized hardware feature vectors and standardized task / instruction feature vectors. Model Inference: The server calculates the corrected matching degree between hardware and tasks through the hardware-task dynamic mapping model of the model building and inference module. The matching degree between GPU nodes and bursty multimodal inference tasks reaches 0.92, the matching degree between NPU nodes and ordinary text generation tasks reaches 0.88, and the matching degree between CPU nodes and ordinary question-answering interaction tasks reaches 0.85. At the same time, the probability of bursty tasks is predicted to be 0.65 through the time-series prediction scheduling model, triggering the buffer resource reservation mechanism to reserve 30% of idle GPU / NPU resources as buffer resources. Initial scheduling decision generation: Based on the matching degree value and prediction results, the server allocates burst tasks to GPU nodes and normal tasks to NPU / CPU nodes, configures register resources for each task (reserving 15% resource redundancy), calculates task execution priorities (burst task priority ≥ 1.5, normal task priority 0.5-1.2), generates initial scheduling decisions and sends them to hardware nodes. Real-time monitoring and dynamic adjustment: The server collects hardware operation data and task execution data at a frequency of 50ms / time, calculates a comprehensive deviation value of 0.12 (≤ preset threshold 0.2), and maintains the initial scheduling decision; when the test reaches 1 hour, the number of sudden tasks increases, the comprehensive deviation value rises to 0.25, the server updates the time-series predictive scheduling model online, re-optimizes the scheduling decision, increases the resource configuration of GPU nodes, adjusts the task execution priority, and ensures the rapid execution of sudden tasks; Fault-tolerant rescheduling test: After 1.5 hours of testing, a CPU node failure was simulated (fault judgment value F_i=1.2≥1.0). The server quickly detected the hardware failure, removed the faulty node, and reassigned the unfinished tasks on the faulty node to the second-highest matching idle CPU node, increasing the task execution priority by 10%. The fault recovery time was ≤200ms, and it did not affect the overall task execution. Test End: After the 2-hour test, the server recorded all task execution data and hardware operation data, and compiled test metrics.

[0066] Compare the testing process of the scheduling method in document 1 The computational resource scheduling method described in Comparison Document 1 is deployed, which divides tasks into abnormal tasks (sudden tasks) and normal tasks, and hardware nodes into normal nodes and buffer nodes. Resources are allocated through node weight (resource processing capacity / number of requests) and number of connections, with priority given to abnormal tasks. During the test, the task and hardware environment are kept consistent with the present invention, and test indicators are recorded.

[0067] Compare the testing process of the scheduling method in document 2. The instruction optimization scheduling method described in Comparison Document 2 is deployed, which converts the large model into a computation task graph, determines the scheduling priority based on the importance and timeliness of the instruction path, and dynamically adjusts the scheduling strategy in combination with the register pressure status; during the test, the task and hardware environment are kept consistent with the present invention, and the test indicators are recorded.

[0068] Test Results and Analysis The test results of the three scheduling methods are shown in Table 5 below. By comparing the test indicators, the practical application advantages of the scheduling method of the present invention are analyzed.

[0069] Table 5. Comparison of test results for the three scheduling methods The test results show that the scheduling method of this invention is significantly superior to the two existing scheduling methods in all core indicators. The specific advantages are analyzed as follows: Significant improvement in hardware resource utilization: The hardware resource utilization rate of this invention reaches 82.5%, which is 41.5% higher than that of Comparative Document 1 and 25.6% higher than that of Comparative Document 2. This is because this invention constructs a hardware-task dynamic mapping model, achieving fine-grained matching between heterogeneous hardware and tasks / instructions. It also considers the underlying characteristics of hardware, such as computing power, registers, and storage, avoiding resource idleness and waste. In contrast, Comparative Document 1 only performs coarse-grained partitioning of hardware into ordinary / buffered nodes, and Comparative Document 2 only adapts to specific homogeneous computing units; neither can achieve deep collaborative adaptation of heterogeneous hardware, resulting in lower resource utilization.

[0070] The task response latency is significantly reduced: the average task response latency of this invention is only 45.2ms, which is 64.8% lower than that of Comparative Document 1 and 53.3% lower than that of Comparative Document 2. This is because this invention constructs a time-series predictive scheduling model, which can predict the probability of sudden tasks in advance and reserve buffer resources, realizing a "proactive prediction-proactive allocation" scheduling mode, avoiding the lag in resource allocation for post-event response; while Comparative Documents 1 and 2 are both passive response scheduling, which require reallocation of resources when facing sudden tasks, resulting in a significant increase in task response latency.

[0071] The scheduling success rate is close to 100%: The scheduling success rate of this invention reaches 99.5%, which is 14.1% higher than that of comparative document 1 and 7.6% higher than that of comparative document 2. The reason is that this invention establishes a complete fault-tolerant rescheduling mechanism, which can quickly detect hardware failures and task execution anomalies, and generate accurate fault-tolerant rescheduling decisions. Fault recovery time is short and has almost no impact on task execution. In contrast, comparative document 1 only temporarily stores tasks in a cache database, and comparative document 2 does not design a fault-tolerant mechanism for hardware failures. When faced with hardware failures and task anomalies, task interruption is likely to occur, resulting in a lower scheduling success rate.

[0072] The register resource conflict rate is extremely low: the register resource conflict rate of this invention is only 1.2%, which is 92.4% lower than that of comparative document 1 and 85.5% lower than that of comparative document 2. The reason is that this invention predicts register resource requirements in advance through a time-series prediction model, reserving reasonable resource redundancy when configuring resources for tasks. At the same time, it dynamically adjusts resource configuration in real time by combining real-time monitoring, thus avoiding competition for register resources. In contrast, comparative document 1 does not consider fine-grained configuration of registers, and comparative document 2 only adjusts its strategy when the register pressure exceeds the threshold, which is prone to register resource conflicts.

[0073] Furthermore, during actual testing, the scheduling system of this invention demonstrated excellent dynamic adaptability and scalability: when a new inference task is added, the server can complete a new round of scheduling within 500ms without manual intervention; when the number of hardware nodes is increased, only hardware identifiers need to be added to the data acquisition module, and the system can automatically complete hardware feature acquisition and matching degree calculation, resulting in low adaptation costs. In contrast, the scheduling systems in comparison documents 1 and 2 require manual adjustment of node weights and scheduling priorities when facing changes in tasks and hardware, exhibiting poor flexibility and scalability.

[0074] Test Conclusion This practical application test demonstrates that the intelligent scheduling method and system based on large-scale artificial intelligence models of this invention can effectively solve the technical problems of insufficient deep collaborative adaptation of hardware resources, passive response to unknown demands, and weak fault tolerance in existing scheduling methods. It achieves fine-grained adaptation of heterogeneous hardware, proactive predictive scheduling of tasks / instructions, dynamic optimization of scheduling decisions, and fault-tolerant rescheduling of hardware / tasks, significantly improving hardware resource utilization, task response efficiency, and scheduling success rate, reducing register resource conflict rate, and demonstrating good dynamic adaptability and scalability. It can meet the real-time intelligent scheduling needs of large-scale model training, inference, and other scenarios, and has significant practical application and promotion value.

[0075] The intelligent scheduling method and system based on artificial intelligence large model of the present invention takes the server as the sole execution entity and follows the core logic of computer data processing, which includes data acquisition, data processing, model inference, decision generation, real-time monitoring, dynamic adjustment and fault-tolerant rescheduling. By constructing a hardware-task dynamic mapping model and a time-series predictive scheduling model, it breaks through the technical barriers of existing large model scheduling technology in terms of heterogeneous hardware adaptation and predictive scheduling, and realizes full-process automation, intelligence and refinement of large model scheduling.

[0076] The scope of protection of this invention is not limited to the descriptions of the above embodiments and usage examples. Conventional modifications, substitutions, and extensions made by those skilled in the art based on the technical solutions of this invention are all within the scope of protection of this invention. Any modifications, equivalent substitutions, and improvements made to this invention without departing from the spirit and essence of the technical solutions of this invention should be included within the scope of protection of this invention.

Claims

1. An intelligent scheduling method based on a large-scale artificial intelligence model, wherein the method uses a server as the execution entity, characterized in that, The method includes: S1: Collect hardware feature data of heterogeneous hardware clusters and task / instruction feature data of large model tasks / instructions to be scheduled, and store the hardware feature data and task / instruction feature data in a local database; S2: Retrieve the hardware feature data and task / instruction feature data from the local database, and perform cleaning, normalization and feature quantization processing on the two types of data respectively to obtain standardized hardware feature vectors and standardized task / instruction feature vectors; S3: Construct a hardware-task dynamic mapping model, input the standardized hardware feature vector and the standardized task / instruction feature vector into the hardware-task dynamic mapping model, and calculate the matching degree value between hardware and task / instruction; Construct a timing prediction scheduling model, input the preprocessed timing monitoring data into the timing prediction scheduling model, and output the resource requirement prediction value, hardware state change prediction value, and probability value of sudden task occurrence for task / instruction. S4: Based on the matching degree value, resource demand prediction value, hardware status change prediction value, and probability value of sudden task occurrence, combined with preset scheduling constraints, an initial intelligent scheduling decision is generated. The initial intelligent scheduling decision includes hardware allocation results, task / instruction execution priority, and resource configuration parameters. S5: Real-time acquisition of hardware operation data and task / instruction execution data during the scheduling and execution process; feature extraction of the data to obtain a real-time monitoring feature vector; calculation of the difference between the real-time monitoring feature vector and the predicted values ​​of hardware state changes and resource requirements to obtain a deviation value. S6: Determine whether the deviation value exceeds the preset deviation threshold. If it does not exceed the threshold, maintain the initial intelligent scheduling decision. If it exceeds the threshold, input the real-time monitoring feature vector into the time-series prediction scheduling model for online update to obtain the updated prediction value. Based on the updated prediction value and the matching degree value, re-optimize the scheduling decision and send it to the heterogeneous hardware cluster for execution. S7: Real-time detection of hardware fault data and abnormal execution data of tasks / instructions in heterogeneous hardware clusters to determine whether there is a hardware fault or abnormal execution of tasks / instructions. If not, the current scheduling decision continues to be executed; if so, the matching degree is recalculated based on the hardware-task dynamic mapping model to generate a fault-tolerant rescheduling decision, replace the faulty hardware allocation result and reallocate the abnormal tasks / instructions.

2. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S1, the hardware feature data includes hardware type dimension data, hardware performance dimension data, and hardware status dimension data. The hardware type dimension data includes the hardware identifier, computing unit type, and register configuration parameters of the CPU / GPU / TPU / NPU. The hardware performance dimension data includes floating-point operation capability, storage bandwidth, network transmission rate, and number of registers. The hardware status dimension data includes CPU utilization, GPU memory usage, register utilization, hardware temperature, and network latency. The task / instruction feature data includes task type dimension data, task attribute dimension data, and instruction feature dimension data. The task type dimension data includes task identifiers for large model training tasks, inference tasks, and data preprocessing tasks. The task attribute dimension data includes computational complexity, data volume, task urgency, task importance, and execution time requirements. Instruction characteristic dimension data includes instruction register requirement, data dependency depth, computation unit affinity, and instruction execution time.

3. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S2, feature quantization processing is performed on the hardware feature data and task / instruction feature data, specifically as follows: The hardware feature data is quantized using a linear weighting method to obtain the quantized hardware feature values. The calculation formula is as follows: in, For the first The characteristic quantization value of each hardware component The number of dimensions of hardware features. For the first The weight coefficients of each hardware feature dimension, and =1, For the first The first hardware Normalized values ​​for each feature dimension; The task / instruction feature data is quantized using a linear weighting method to obtain the quantized value of the task / instruction feature. The calculation formula is as follows: in, For the first The characteristic quantization value of each task / instruction The number of task / instruction feature dimensions. For the first The weight coefficients of each task / instruction feature dimension, and =1, For the first The first task / instruction Normalized values ​​for each feature dimension; The quantized values ​​of the hardware features and the quantized values ​​of the task / instruction features are subjected to vector normalization to obtain standardized hardware feature vectors and standardized task / instruction feature vectors. The normalization formula is as follows: ， in, For the first The standardized hardware feature vector of a hardware unit. For the first A standardized task / instruction feature vector for each task / instruction. , These are the L2 norms of the original hardware feature vector and the original task / instruction feature vector, respectively.

4. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S3, a hardware-task dynamic mapping model is constructed, and the matching degree value between hardware and task / instruction is calculated, specifically as follows: A matching degree calculation module for the hardware-task dynamic mapping model is constructed using the cosine similarity algorithm. The calculation formula is as follows: in, For the first The hardware and the first The matching degree value of each task / instruction is set, with a value ranging from [0,1], where a larger value indicates a higher matching degree. The matching degree value is then corrected using a collaborative filtering algorithm to obtain the corrected matching degree value. The calculation formula is: in, This is a correction factor, with a value range of [0, 0.5]. In order to be with the first The number of hardware devices of the same type For the first The hardware and the first Similarity values ​​for similar hardware.

5. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S3, the time-series predictive scheduling model is constructed as follows: A time-series predictive scheduling model is constructed using an LSTM neural network combined with an attention mechanism. The input to the model is the feature matrix of the time-series monitoring data. ,in for The monitoring feature vector at any given time; The hidden layer output of the LSTM is weighted using an attention mechanism. The formula for calculating the attention weights is as follows: in, for Time Hidden Layer Output Attention weights For trainable weight matrix, This is the bias term; the model's fusion output is: fused output The input is fed into the fully connected layer to obtain the prediction result. The loss function of the model is the mean squared error loss function, and the calculation formula is as follows: Where N is the number of training samples, For the true value, These are the model's predicted values.

6. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S4, the initial intelligent scheduling decision is generated, specifically as follows: Based on the probability value P of the occurrence of the unexpected task, determine whether there is a risk of an unexpected task. ( To preset a sudden risk threshold, a preset proportion of hardware resources are reserved as buffer resources. The calculation formula is: Based on the corrected matching degree value Perform an initial matching of hardware with tasks / instructions, and allocate hardware in descending order of matching score; Based on the predicted resource requirements, the resources are configured for the matched task / instruction, including the number of register resources configured. The calculation formula is: in For the first Number of base registers required for each task / instruction Forecast values ​​of register resource requirements, This is a resource reservation coefficient, with a value range of [0, 0.2]. Quantification value based on task / instruction characteristics Calculate execution priority The calculation formula is: in, This is the priority weight coefficient, and , For the first The maximum corrected match value for a task / instruction.

7. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S5, calculating the deviation value specifically involves: calculating the hardware state deviation value respectively. Deviation value of resource demand The calculation formula is: in, For the number of hardware, For the first The actual state value of each piece of hardware. For the first The state prediction value of each hardware component. For the first The nominal status value of each hardware component; For the number of tasks / instructions, For the first The actual resource requirements of each task / instruction. For the first Forecasted resource requirements for each task / instruction. For the first The nominal value of resource requirements for each task / instruction; Calculate the overall deviation value The calculation formula is: in, This is the weighting coefficient for hardware status deviation, with a value range of [0.4, 0.6].

8. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S7, hardware fault data and abnormal execution data of tasks / instructions are detected in real time. Specifically, a threshold detection method combined with trend analysis is used to detect hardware faults, and a hardware fault judgment value is determined. The calculation formula is: in, For hardware fault monitoring indicators, For the first The first hardware The actual value of each monitoring indicator For the first Normal thresholds for each monitoring indicator This is the trend coefficient. The rate of change of hardware status indicators; like ( (If the preset fault determination threshold is used), then the fault determination threshold is determined. A hardware component malfunctioned; Execution schedule deviation detection is used to identify task / instruction execution anomalies, and the anomaly judgment value is... The calculation formula is: in, For the first The actual execution progress of each task / instruction. This refers to the actual execution time. To preset the execution schedule, The preset execution time; like ( (If the preset anomaly detection threshold is used), then the first anomaly is determined. An error occurred during the execution of a task / instruction.

9. The intelligent scheduling method based on a large artificial intelligence model according to claim 1, characterized in that, In step S7, generating a fault-tolerant rescheduling decision specifically involves: if a hardware failure is detected, removing the faulty hardware and selecting a corrected matching degree value from the hardware cluster. The second-highest priority hardware that is idle is used as a replacement hardware to reallocate unfinished tasks / instructions on the faulty hardware; if an abnormal task / instruction execution is detected, the resource requirement deviation value of the abnormal task / instruction is calculated. If the deviation value is caused by insufficient resources, the resource allocation is increased based on the updated resource requirement prediction value; if the deviation value is caused by low hardware matching degree, the hardware-task matching process is re-executed to select hardware with a higher matching degree for reallocation.

10. An intelligent scheduling system based on a large-scale artificial intelligence model, characterized in that, The system is deployed on a server and includes: The data acquisition module is used to collect hardware feature data of heterogeneous hardware clusters and task / instruction feature data of large model tasks / instructions to be scheduled, and store the hardware feature data and task / instruction feature data in a local database. The data preprocessing module is used to retrieve the hardware feature data and task / instruction feature data from the local database, and perform cleaning, normalization and feature quantization on the two types of data to obtain standardized hardware feature vectors and standardized task / instruction feature vectors. The model building and inference module is used to build a hardware-task dynamic mapping model, calculate the matching degree between hardware and tasks / instructions, and build a timing prediction scheduling model to output the predicted resource requirements, hardware state changes, and probability of sudden tasks for tasks / instructions. The scheduling decision generation module is used to generate an initial intelligent scheduling decision based on the matching degree value, the predicted value of resource demand, the predicted value of hardware status change, and the probability value of sudden task occurrence, combined with preset scheduling constraints. The real-time monitoring module is used to collect hardware operation data and task / instruction execution data in real time during the scheduling process, extract real-time monitoring feature vectors, calculate deviation values, and determine whether they exceed the preset deviation threshold. The dynamic adjustment module is used to update the time-series prediction scheduling model online if the deviation value exceeds the preset deviation threshold, and then re-optimize the scheduling decision based on the updated prediction value and matching degree value before issuing it for execution. The fault-tolerant rescheduling module is used to detect hardware fault data and abnormal execution data of tasks / instructions in real time. If a fault or abnormality is found, it generates a fault-tolerant rescheduling decision and executes it.

Images