A dynamic node allocation system based on a GPT model and a method thereof

Abstract

This invention discloses a dynamic node allocation system and method based on the GPT model, relating to the fields of artificial intelligence and distributed systems. The method includes: determining node feature dimensions based on task requirements and generating token identifiers through encoding; combining these to generate a task identifier sequence; generating temporal feature vectors based on timestamp data and historical data; obtaining fused features through dual attention branches and interaction matrices, and calculating the original node quantity score, the original node fit score, and the cross-domain coupling potential factor to generate the optimal number of nodes and node identifiers; verifying node availability and replacing unavailable nodes to obtain the allocation result; executing according to the result and providing status feedback. By fusing task and temporal features through branch attention and interaction matrix modeling, combined with adaptive temperature calibration and availability replacement feedback mechanisms, controllable and stable allocation of node quantity and node selection is achieved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and distributed systems, specifically to a dynamic node allocation system and method based on the GPT model. Background Technology

[0002] With the development of distributed computing technology, more and more tasks need to be executed in multi-node environments, and the rationality of node allocation directly affects task processing efficiency and resource consumption. In scenarios such as cloud computing and edge computing, how to select an appropriate number of nodes and match the optimal nodes according to task requirements has become a key issue in improving system performance.

[0003] In existing technologies, node allocation typically relies on static configuration strategies, rule engine strategies, or load balancing strategies to map tasks to nodes. Static configuration strategies tend to pre-determine the number and set of nodes according to task categories, rule engine strategies tend to filter node resource status based on threshold conditions, and load balancing strategies tend to distribute traffic or make weighted selections based on current load metrics. While these solutions are easy to deploy in engineering implementation, they usually rely on local metrics or fixed rules, making it difficult to uniformly model the high-dimensional relationship between task requirements and node attributes, and also making it difficult to maintain the ability to continuously adjust the allocation strategy when tasks reach rhythmic changes.

[0004] Existing technologies typically face several problems: insufficient flexibility, as static node allocation strategies cannot dynamically adjust the number of nodes according to the specific needs of a task. For example, a simple task that can be efficiently completed with only one node may be assigned multiple nodes, resulting in wasted resources. Complex tasks may suffer from low execution efficiency due to insufficient allocated nodes, because the strategy does not consider the differences and dynamic changes in task requirements. Low matching accuracy, as load balancing algorithms only focus on node load status and ignore the compatibility between task requirements and node attributes. For example, tasks with specific resource capacity requirements may be assigned to nodes with insufficient resources, leading to task failure or decreased efficiency. The core reason is the lack of comprehensive analysis of the matching relationship between task characteristics and node attributes. Poor decision determinism, as some dynamic allocation schemes do not introduce effective probability calibration mechanisms, resulting in randomness in the selection of the number of nodes and specific nodes, making it difficult to guarantee the stability of task execution. This is because there is a lack of precise means to control the prediction results.

[0005] To address the aforementioned issues, this invention proposes a dynamic node allocation system and method based on the GPT model, which achieves precise matching between task requirements and node resources, dynamically determines the optimal number of nodes and specific nodes, thereby improving resource utilization and task execution efficiency while ensuring the stability and reliability of allocation decisions. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a dynamic node allocation system and method based on the GPT model to solve the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a dynamic node allocation method based on the GPT model, comprising: Based on the task requirements, obtain the node feature dimensions, encode the feature values ​​of the node feature dimensions into token identifiers, and generate a task identifier sequence. Extract time-series feature vectors based on task timestamp data and historical execution record data; A fusion sequence representation is generated based on the task identifier sequence and temporal feature vector. Fusion features are generated through dual attention branches and cross-branch interaction matrices. The original score of the number of nodes, the original score of node adaptation, and the cross-domain coupling potential factor are calculated. Based on the original scores for the number of nodes, the original scores for node adaptation, and the cross-domain coupling potential factor, the optimal number of nodes and node identifiers are generated. Availability is verified based on the optimal number of nodes and node identifiers, and unavailable nodes are replaced to obtain the node allocation result. Execute tasks based on node allocation results, monitor task execution progress in real time, and collect execution status for feedback.

[0008] The present invention is further configured such that the generated task identifier sequence includes: Analyze the task requirements to obtain the necessary node characteristic dimensions, including node type requirements, load capacity requirements, operating area requirements, resource capacity requirements, and node status requirements. For each node's feature dimension, a set of unique feature values ​​of the same number is pre-defined, and an encoding mapping from feature values ​​to token identifiers is established; Based on the encoding mapping, the feature value corresponding to the feature dimension of each node is converted into a token identifier; The token identifiers are combined according to the preset arrangement order of the node feature dimensions to generate a task identifier sequence.

[0009] The present invention is further configured such that the extraction of time-series feature vectors based on task timestamp data and historical execution record data includes: Task timestamp data includes the current task submission time and the execution time of the most recent similar task; Historical execution record data includes the number of times the same type of task was executed within a preset time window and the historical average execution time interval; Calculate the time interval feature based on the current task submission time and the execution time of the most recent similar task, and combine the time interval feature, the number of similar tasks executed within a preset time window and the historical average execution time interval to generate a time series feature vector; Normalization is performed on the temporal feature vector to obtain the three-dimensional temporal feature vector.

[0010] The present invention is further configured such that generating the fused sequence representation based on the task identifier sequence and the temporal feature vector includes: A task token embedding sequence is generated based on the task identifier sequence, and a time-series token embedding sequence is generated based on the time-series feature vector. A fusion sequence is generated by concatenating the task token embedding sequence and the time sequence token embedding sequence in a preset order; A position embedding is generated based on the token position of the fused sequence, and the position embedding is injected into the fused sequence to obtain the fused sequence representation.

[0011] The present invention is further configured such that generating fused features through dual attention branches and cross-branch interaction matrices includes: Based on the fused sequence representation, query vectors, key vectors, and value vectors are generated through shared projection; In the task feature branch, generate a task branch attention score matrix within the task token range based on the query vector and key vector; In the temporal feature branch, a temporal branch attention score matrix is ​​generated within the temporal token range based on the query vector and key vector. Fill the intersection area of ​​task tokens and time-series tokens according to the cross-branch interaction matrix, and generate a fusion attention matrix by combining the task branch attention score matrix and the time-series branch attention score matrix; The fused features are obtained by performing weighted convergence on the value vectors based on the fusion attention matrix.

[0012] The present invention is further configured such that the original score of the number of computing nodes, the original score of node adaptation, and the cross-domain coupling potential factor include: Based on the fusion features, the original score for the number of nodes is generated through the node number output header, and the original score for node adaptation is generated through the node adaptation output header. Generate task domain energy terms based on task branch attention score matrix, generate time domain energy terms based on time branch attention score matrix, and generate cross-domain energy terms based on cross-branch interaction matrix; The cross-domain coupling potential energy factor is generated based on the task domain energy term, the time-series domain energy term, and the cross-domain energy term.

[0013] The present invention is further configured such that generating the optimal number of nodes and node identifiers includes: Calculate the adaptive temperature parameters based on the cross-domain coupling potential energy factor; Based on the adaptive temperature parameters, the original fraction of the number of nodes is subjected to temperature scaling and probabilistic transformation to obtain the probability distribution of the number of nodes, and the number of nodes with the highest probability is selected as the optimal number of nodes. The node adaptation probability distribution is obtained by performing temperature scaling and probabilization on the original node adaptation scores based on the adaptive temperature parameters. Based on the optimal number of nodes, the node adaptation probability distribution is sorted by probability magnitude, and the nodes that are ranked first and whose number is equal to the optimal number of nodes are selected.

[0014] The present invention is further configured such that obtaining the node allocation result includes: Node resource status data is obtained based on the node identifier. The node resource status data includes node type, current load, region, resource capacity, and operating status. Based on the node resource status data, perform availability checks on the nodes corresponding to the node identifiers and generate a set of unavailable node identifiers. When the set of unavailable node identifiers is not empty, candidate replacement node identifiers are selected according to the node adaptation probability distribution, and node replacement is performed to obtain the set of replaced node identifiers; The node allocation result is generated based on the replaced set of node identifiers and the optimal number of nodes.

[0015] The present invention is further configured such that executing tasks according to node allocation results, monitoring task execution progress in real time, and collecting execution status feedback include: Based on the node allocation results, the task will be distributed to the node corresponding to the node allocation result. Collect task execution progress data based on the execution process after the task is issued; Collect execution status data based on task execution progress data; The execution status data is written to the node resource status data for feedback.

[0016] This invention also provides a dynamic node allocation system based on the GPT model, the system comprising: The task feature encoding module obtains node feature dimensions according to task requirements and encodes the feature values ​​of the node feature dimensions into token identifiers to generate a task identifier sequence. The time-series feature encoding module extracts time-series feature vectors based on task timestamp data and historical execution record data; The GPT processing module generates a fused sequence representation based on the task identifier sequence and temporal feature vector, generates fused features through dual attention branches and cross-branch interaction matrices, and calculates the original score of the number of nodes, the original score of node adaptation, and the cross-domain coupling potential factor. The temperature scaling decision module generates the optimal number of nodes and node identifiers based on the original scores of the number of nodes, the original scores of node adaptation, and the cross-domain coupling potential energy factor. The node resource management module performs availability checks based on the optimal number of nodes and node identifiers, replaces unavailable nodes, and generates node allocation results. The results output and execution module executes tasks based on the node allocation results, monitors the task execution progress in real time, and feeds back the execution status to the node resource management module.

[0017] This invention provides a dynamic node allocation system and method based on the GPT model. The method obtains node feature dimensions according to task requirements, encodes the feature values ​​of these dimensions into token identifiers to generate a task identifier sequence; extracts temporal feature vectors based on task timestamp data and historical execution record data; generates a fused sequence representation based on the task identifier sequence and temporal feature vectors, generates fused features through dual attention branches and cross-branch interaction matrices, and calculates the original node quantity score, original node adaptation score, and cross-domain coupling potential factor; generates an optimal number of nodes and node identifiers based on these scores; performs availability verification based on the optimal number of nodes and node identifiers, replaces unavailable nodes to generate node allocation results; executes tasks based on the node allocation results, monitors task execution progress in real time, and collects execution status feedback. The beneficial effects include: 1. Cross-domain fusion modeling improves allocation consistency: Based on task identifier sequence and temporal feature vector, a fusion sequence representation is constructed. Fusion features are generated through task branch attention, temporal branch attention and cross-branch interaction matrix. At the same time, the original scores of node number and node adaptation are output, so that the determination of node number and node identifier selection share the same fusion basis, reducing the inconsistency caused by the allocation strategy fluctuating with the input. 2. Adaptive temperature calibration enhances decision controllability: The cross-domain coupling potential energy factor is introduced to calculate the adaptive temperature parameter. Temperature scaling and probabilization are performed on the original score of the number of nodes and the original score of node adaptation, making the sharpness of the probability distribution adjustable, taking into account both determinism and flexibility, avoiding decision rigidity caused by fixed thresholds or single rules, and improving the controllability of the number of nodes and node identification output. 3. Availability verification and replacement ensure an executable closed loop: After generating the optimal number of nodes and node identifiers, availability verification is performed based on node resource status data. Unavailable nodes are replaced to obtain the node allocation results. During task execution, progress is monitored in real time, execution status is collected, and node resource status data is updated and fed back, realizing a closed loop of allocation-execution-feedback, improving the executability and continuous adaptability of the allocation results.

[0018] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0019] 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 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. In the drawings: Figure 1 A flowchart illustrating a dynamic node allocation method based on the GPT model, as shown in an exemplary embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the structure of a dynamic node allocation system based on the GPT model, as an exemplary embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0021] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0022] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0023] Example 1: A dynamic node allocation method based on the GPT model, such as Figure 1 As shown, it includes: Based on the task requirements, obtain the node feature dimensions, encode the feature values ​​of the node feature dimensions into token identifiers, and generate a task identifier sequence. Extract time-series feature vectors based on task timestamp data and historical execution record data; A fusion sequence representation is generated based on the task identifier sequence and temporal feature vector. Fusion features are generated through dual attention branches and cross-branch interaction matrices. The original score of the number of nodes, the original score of node adaptation, and the cross-domain coupling potential factor are calculated. Based on the original scores for the number of nodes, the original scores for node adaptation, and the cross-domain coupling potential factor, the optimal number of nodes and node identifiers are generated. Availability is verified based on the optimal number of nodes and node identifiers, and unavailable nodes are replaced to obtain the node allocation result. Execute tasks based on node allocation results, monitor task execution progress in real time, and collect execution status for feedback.

[0024] The present invention is further configured such that the generated task identifier sequence includes: The process analyzes task requirements to obtain the necessary node feature dimensions, including node type requirements, load capacity requirements, operating region requirements, resource capacity requirements, and node status requirements. Specifically, upon receiving a task creation request, the task's requirement description fields are read, including task type, data processing scale, geographical constraints, latency constraints, availability constraints, and resource preferences. Based on preset feature dimension selection rules, the dimension entries directly related to node selection are located in the task requirement description, generating a node feature dimension list. This list includes node type requirements, load capacity requirements, operating region requirements, resource capacity requirements, and node status requirements. Node type requirements are inferred from the task type or task interface type field; load capacity requirements are inferred from the task's pre-calculation power consumption or concurrency level field; operating region requirements are extracted from the geographical constraint field; resource capacity requirements are extracted from the combined constraints of memory, storage, and bandwidth requirements; and node status requirements are extracted from the availability constraint field to allow only nodes in normal state or nodes in degraded state. After completion, a node feature dimension list arranged in a fixed order is obtained. For each node feature dimension, a pre-defined set of the same number of unique feature values ​​is established, and an encoding mapping from feature values ​​to token identifiers is created. Specifically, feature value dictionaries are established for node type requirements, load capacity requirements, operating area requirements, resource capacity requirements, and node status requirements, with each dictionary pre-defined with the same number of feature value entries. Each entry is a unique feature value, constrained by the principle of "no repetition within the same dimension and no reuse between different dimensions" to avoid semantic aliasing caused by the same identifier appearing in different dimensions. Token identifiers are assigned to the unique feature value set for each dimension, forming a one-to-one correspondence between "feature values ​​and token identifiers" and written into the encoding mapping table. The encoding mapping table contains the dimension identifier, unique feature value text, token identifier, version number, and effective time. The version number is used to maintain the traceability of historical tasks when the dictionary is updated, and the effective time is used to ensure that the same batch of tasks uses the same mapping version. The encoding mapping table is loaded into the memory index structure of the task feature encoding module, with the index key being "dimension identifier plus unique feature value" and the index value being the token identifier. The feature values ​​corresponding to each node's feature dimension are converted into token identifiers based on the encoding mapping. Specifically, for each node's feature dimension, a unique feature value is first determined according to the task requirements. This is done by selecting a value range matching the dimension from the task requirement field or enumerating values, and then mapping it to a preset unique feature value according to the dictionary's normalization rules. The normalization rules include: region names are standardized to standard area codes, load levels are standardized to preset load tiers, resource capacity is standardized to preset capacity tiers, and node status is standardized to a preset status set. When task requirements are non-standard, a rule-based transformation is performed before mapping to a unique feature value. Then, the token identifier is searched in the encoding mapping table using "dimension identifier plus unique feature value" as the key. If the search is successful, the corresponding token identifier is output; otherwise, the default token identifier for that dimension is output according to the reserved entry in the mapping table, and the reason for the error is recorded for subsequent dictionary completion. After completion, a set of token identifiers corresponding one-to-one with each node's feature dimension is obtained. Based on a preset order of node feature dimensions, token identifiers are combined to generate a task identifier sequence. Specifically, token identifiers from each dimension are combined into a task identifier sequence using a fixed order consistent with the node feature dimension list: node type requirement, load capacity requirement, operating area requirement, resource capacity requirement, and node status requirement. During combination, the sequence position is kept semantically linked to the dimension; the first position in the sequence carries only the node type token identifier, the second position carries only the load capacity token identifier, and so on, ensuring that subsequent model processing can stably identify the semantic source of each position.

[0025] The present invention is further configured such that the extraction of time-series feature vectors based on task timestamp data and historical execution record data includes: The task timestamp data includes the current task submission time and the execution time of the most recent similar task. Specifically, it receives event records of the current task entering the scheduling queue, reads the submission time field from the event record as the current task submission time, and reads the similarity identification information of the task as the filtering condition for similar tasks. The similarity identification information includes at least a task type identifier, a business scenario identifier, or a task template identifier. Based on the similarity identification information, it retrieves the most recent "execution completed" record of similar tasks from the historical task index. During the retrieval, the record status is limited to "completed" and the record time is earlier than the current task submission time. When multiple records that meet the conditions are retrieved, the completion time of the first record in reverse order of completion time is selected as the most recent similar task execution time. When no record that meets the conditions is retrieved, the most recent similar task execution time is set to the default status, and the default reason identifier is recorded. Historical execution record data includes the number of times similar tasks were executed within a preset time window and the historical average execution time interval. Specifically, a query condition with similarity discrimination information as the key is established in the historical execution record database, and the window length and window endpoint alignment rules are set for the preset time window. The window length can be selected as 24 hours, and the window endpoint alignment rule can be selected as looking back from the right endpoint of the current task submission time, or aligning with the window endpoint at midnight of a natural day. The execution completion records of similar tasks within the preset time window are retrieved according to the query conditions, and the number of records retrieved is determined as the number of times similar tasks were executed within the preset time window. Regarding the determination of the historical average execution time interval, the completion time sequence of similar tasks within the preset time window is first sorted in ascending order by time, and then the interval between two adjacent completion times is formed into an interval sequence. When there are valid entries in the interval sequence, the arithmetic mean of the interval sequence is used as the historical average execution time interval. When the interval sequence is empty or there are missing timestamps that prevent the formation of valid interval entries, the historical average execution time interval is set to the default state. The time interval feature is calculated based on the current task submission time and the execution time of the most recent similar task. This time interval feature, the number of similar tasks executed within a preset time window, and the historical average execution time interval are combined to generate a time-series feature vector. Specifically, when the execution time of the most recent similar task is valid, the time interval feature is determined based on the chronological relationship between the current task submission time and the most recent similar task execution time, with the time interval feature representing the time span between the two. When the execution time of the most recent similar task is in a default state, the time interval feature adopts a default value strategy. The default value can be the preset time window length or a default alternative value of the historical average execution time interval. The default value strategy is associated with the default reason identifier. Subsequently, the time interval feature, the number of similar tasks executed within the preset time window, and the historical average execution time interval are combined in a fixed order to form the time-series feature vector. The combination order is fixed as follows: time interval feature first, execution count second, and historical average execution time interval last. The time-series feature vector, along with the current task identifier, similarity discrimination information, and preset time window configuration, are written into the time-series input data structure. Normalization is performed on the temporal feature vector to obtain a three-dimensional temporal feature vector. Specifically, to ensure consistency of dimensions for different temporal features, a normalization benchmark set is established. This set maintains a minimum and maximum benchmark value for each dimension of the temporal feature. The minimum and maximum benchmark values ​​can be obtained from historical rolling statistics. The statistical range of the rolling statistics can be set to a sample set of similar tasks from the most recent 30 days or the most recent 100 days. The update cycle can be set to daily or hourly updates, and the statistical version number is recorded with each update. During normalization, minimum and maximum scaling are applied to the time interval feature, execution count, and historical average execution time interval. The scaling rule is to use the difference between the current feature value and the minimum benchmark value as a relative quantity, and then use the difference between the maximum and minimum benchmark values ​​as a scale quantity for mapping. When the scale quantity is zero or the feature value is in a default state, a protection strategy is adopted to output a preset median or a zero value, and a normalization protection flag is recorded. After completing the dimensional scaling of the three-dimensional features, the three-dimensional temporal feature vector is output.

[0026] The present invention is further configured such that generating the fused sequence representation based on the task identifier sequence and the temporal feature vector includes: A task token embedding sequence is generated based on the task identifier sequence, and a time-series token embedding sequence is generated based on the time-series feature vector. Specifically, each token identifier in the task identifier sequence is used as an index to retrieve the embedding vector corresponding to that token identifier from a pre-set token embedding table. These vectors are then arranged sequentially according to the order of the task identifier sequence to form the task token embedding sequence. The token embedding table is loaded during system initialization, and the embedding vector dimension is a preset embedding dimension. The values ​​of the embedding vectors are learned during the model training phase and remain fixed during the inference phase. Subsequently, the three-dimensional time-series feature vector is input into the time-series embedding generator for transformation. The time-series embedding generator uses a combination of linear mapping and normalization to transform the three-dimensional numerical vector into a time-series embedding vector with the same dimension as the token embedding vector. The time-series embedding vector is then split into several time-series token embedding vectors according to a preset time-series token splitting rule. The splitting rule includes the number of splits, the splitting order, and the splitting position. The number of splits is consistent with the number of time-series tokens configured on the model input side, and the splitting order is consistent with the order of the three-dimensional time-series features. The split time-series token embedding vectors are then arranged in the splitting order to form the time-series token embedding sequence. A fused sequence is generated by concatenating the task token embedding sequence and the time-series token embedding sequence in a preset order. Specifically, a sequence concatenation operation is performed on the task token embedding sequence and the time-series token embedding sequence. The concatenation order adopts a fixed configuration and remains unchanged during the model inference stage. The fixed configuration includes the task token embedding sequence first and the time-series token embedding sequence second. During concatenation, the internal order of the task token embedding sequence is kept consistent with the task identifier sequence, and the internal order of the time-series token embedding sequence is kept consistent with the time-series token splitting order. After concatenation, the fused sequence is obtained. The length of the fused sequence is determined by the number of task tokens and the number of time-series tokens. The token type at each position in the sequence is fixed by the position interval rule, that is, the previous position interval corresponds to the task token, and the next position interval corresponds to the time-series token. The fused sequence, the number of task tokens, the number of time-series tokens, and the concatenation configuration identifier are written into the fused input data structure. A position embedding is generated based on the token position of the fused sequence. This position embedding is then injected into the fused sequence to obtain a fused sequence representation. Specifically, a position index sequence is generated based on the sequence length of the fused sequence. The position index sequence increments sequentially from the starting position, generating a unique position index for each token position in the fused sequence. The position embedding vector is retrieved from a pre-set position embedding table based on the position index sequence. This table is loaded during system initialization, and the dimension of the position embedding vector is consistent with the dimension of the token embedding vector. The values ​​of the position embedding vectors are learned during the training phase or generated and fixed by a preset position encoding rule. The token embedding vector at each position is then element-wise superimposed with the position embedding vector at that position to obtain the fused representation vector for that position. These vectors are arranged in positional order to form the fused sequence representation. To avoid position inconsistencies caused by changes in sequence length, the sequence length, position embedding table version number, and position index generation rule identifier are simultaneously written into the fused sequence representation generation record. When the sequence length exceeds the coverage of the position embedding table, a preset expansion strategy is used to generate a new position embedding vector and write it into the expansion record.

[0027] The present invention is further configured such that generating fused features through dual attention branches and cross-branch interaction matrices includes: Based on the fused sequence representation, query vectors, key vectors, and value vectors are generated through shared projection. Specifically, the fused sequence representation is input into a shared linear layer, which contains three sets of projection weights, corresponding to query vector generation, key vector generation, and value vector generation, respectively. These three sets of projection weights are learned and determined during model training and remain fixed during inference. A linear mapping is performed on each token position vector in the fused sequence representation to obtain the query vector, key vector, and value vector at that position. To meet the parallel computation requirements of multi-head attention, the query vector, key vector, and value vector are split into several sub-vector groups according to a preset number of heads. Each sub-vector group corresponds to one attention head. The splitting rules include head number configuration, sub-vector dimension configuration, and splitting order configuration, while maintaining the token position order after splitting. In the task feature branch, a task branch attention score matrix is ​​generated within the task token range based on the query vector and key vector. Specifically, the task token range is determined according to the token position interval rules of the fusion sequence, and the task token range corresponds to the previous position interval in the fusion sequence. Based on the task token range, sub-vector groups at corresponding positions are extracted from the split query vector sequence and key vector sequence to form the task branch input. Attention score calculation is performed on the task branch input: taking the query sub-vector group at each task token position as a reference, a similarity score is calculated with the key sub-vector groups at all task token positions to obtain the task branch attention score matrix; the similarity score uses dot product similarity and combines a preset scaling factor to suppress numerical inflation caused by dimensionality increase. In the temporal feature branch, a temporal branch attention score matrix is ​​generated within the temporal token range based on the query vector and key vector. Specifically, the temporal token range is determined according to the token position interval rules of the fused sequence, and the temporal token range corresponds to the latter position interval in the fused sequence. Based on the temporal token range, sub-vector groups at corresponding positions are extracted from the split query vector sequence and key vector sequence to form the temporal branch input. Attention score calculation is performed on the temporal branch input: taking the query sub-vector group at each temporal token position as a reference, similarity scores are calculated with the key sub-vector groups at all temporal token positions to obtain the temporal branch attention score matrix; the similarity scores also use dot product similarity combined with a preset scaling factor. The cross-branch interaction matrix is ​​used to fill the intersection area of ​​task tokens and time-series tokens. A fusion attention matrix is ​​then generated by combining the task branch attention score matrix and the time-series branch attention score matrix. Specifically, a cross-branch interaction matrix is ​​pre-set for the number of task tokens and the number of time-series tokens. The number of rows and columns in the cross-branch interaction matrix are the same as the number of task tokens. The matrix elements are learnable weights, updated during training and fixed during inference. When constructing the fusion attention matrix, the task branch attention score matrix is ​​used as the upper left region, the time-series branch attention score matrix as the lower right region, the cross-branch interaction matrix as the upper right intersection region, and the transpose of the cross-branch interaction matrix as the lower left intersection region, thus forming a complete fusion attention matrix covering the position ranges of both task tokens and time-series tokens. To ensure consistent row normalization in the fusion attention matrix, normalization recalibration is performed on each row to ensure that the weights in each row satisfy the same normalization constraint, avoiding inconsistent weight scales caused by filling the intersection area. Normalization recalibration is performed on all elements in the row, and the normalization strategy identifier is recorded. The fused features are obtained by weighted aggregation of the value vectors based on the fused attention matrix. Specifically, the token position order of the value vector sequence is kept consistent with the fused sequence. For each row of the fused attention matrix, the value vector sequence is weighted and aggregated according to the token position corresponding to the weight, resulting in the output vector corresponding to that token position. The output vectors at each position are arranged in the order of the token positions to obtain the attention output sequence. For multi-head attention, the attention output sequences corresponding to each attention head are concatenated according to a preset concatenation order to obtain the concatenated attention output representation. An output projection transformation is performed on the concatenated attention output representation to make the output representation dimension consistent with the fused sequence representation dimension. Then, a residual connection is performed between the output projection transformation result and the fused sequence representation, and normalization is performed to obtain the first intermediate representation. The first intermediate representation is input into a feedforward neural network for nonlinear transformation. The feedforward neural network adopts a combination structure of two-level linear transformation and activation function. A residual connection is performed between the output of the feedforward neural network and the first intermediate representation, and normalization is performed to obtain the fused features. The fused features are written into the intermediate output structure of the model, and the task branch attention score matrix, temporal branch attention score matrix, cross-branch interaction matrix identifiers are associated with and stored with the fused feature generation record.

[0028] The present invention is further configured such that the original score of the number of computing nodes, the original score of node adaptation, and the cross-domain coupling potential factor include: Based on the fusion features, a raw score for the number of nodes is generated through the node count output header, and a raw score for node adaptation is generated through the node adaptation output header. Specifically, a final normalization process is performed on the fusion features to unify the numerical distribution range of different batches of input. The normalization process adopts a position-by-position normalization method, performing mean-based bias removal and variance scaling on each token position vector to avoid local extrema causing bias in subsequent outputs. The normalized fusion features are input into the node count output header and the node adaptation output header, respectively. The node count output header uses a linear mapping structure to map the fusion features to the node count candidate category space, outputting the raw score sequence corresponding to each node count candidate category. The node count candidate category space is defined by a preset node count set, for example, forming a candidate category set from a preset minimum node count to a maximum node count, and the candidate category order is fixed in the system configuration. The node adaptation output header uses a linear mapping structure to map the fusion features to the candidate node set space, outputting the raw score sequence corresponding to each node in the candidate node set. The candidate node set is provided by the node resource management side, and the node identifier order of the candidate node set remains fixed during this inference and corresponds one-to-one with the output score sequence index. Write the original score of the number of nodes and the original score of node adaptation into the intermediate output data structure, and record the candidate category set version number of the number of nodes, the candidate node set version number and the output order configuration identifier. The task domain energy term is generated based on the task branch attention score matrix, the time-series domain energy term is generated based on the time-series branch attention score matrix, and the cross-domain energy term is generated based on the cross-branch interaction matrix. Specifically, the task branch attention score matrix is ​​first numerically stabilized, including filling missing terms with zeros and truncating abnormally large values ​​with truncation markers. Then, each element in the matrix is ​​squared and summed to obtain the total amount reflecting the internal correlation strength of the task domain, which is used as the task domain energy term. Using the same processing flow as the task domain, the time-series branch attention score matrix is ​​stabilized, and the matrix elements are squared and summed to obtain the total amount reflecting the internal correlation strength of the time-series domain, which is used as the time-series domain energy term. After performing the same stabilization process on the cross-branch interaction matrix, the matrix elements are squared and summed to obtain the total interaction amount. At the same time, the total interaction amount corresponding to the transpose of the cross-branch interaction matrix is ​​summed in the same way. The two parts of the total interaction amount are merged to obtain the cross-domain energy item. The merging method adopts the summation method and records the merging rule identifier to ensure that the cross-domain energy item has a consistent interpretation under different implementations. The task domain energy item, time-series domain energy item, and cross-domain energy item are written into the energy item data structure and associated with the current attention head number configuration identifier. When using multi-head attention, the energy item can be generated separately according to the attention head and then aggregated. The aggregation rule adopts the summation of the energy items of each head to obtain the global energy item, and at the same time, the strategy identifier of whether to aggregate by head is recorded. A cross-domain coupling potential energy factor is generated based on the task domain energy term, the temporal domain energy term, and the cross-domain energy term. Specifically, a ratio-based processing is performed on the task domain energy term, the temporal domain energy term, and the cross-domain energy term to generate the cross-domain coupling potential energy factor. The ratio-based processing adopts a percentage calculation method with the cross-domain energy term as the numerator and the sum of the task domain energy term, the temporal domain energy term, and the cross-domain energy term as the denominator. A preset minimum protection value is added to the denominator to prevent the denominator from being zero. The minimum protection value is a system preset constant, written into the configuration, and fixed during the inference phase. The cross-domain coupling potential energy factor is output as a normalized scalar, and its value increases as the proportion of the cross-domain energy term increases and decreases as the proportion of the cross-domain energy term decreases, used to characterize the weight proportion of cross-domain interaction in the overall association.

[0029] The present invention is further configured such that generating the optimal number of nodes and node identifiers includes: Adaptive temperature parameters are calculated based on the cross-domain coupling potential energy factor. Specifically, temperature parameter value boundaries and temperature adjustment rules are preset. The temperature parameter value boundaries include minimum and maximum temperature values. The temperature adjustment rules include a monotonic mapping direction and sensitivity configuration. The monotonic mapping direction specifies that the temperature parameter decreases when the cross-domain coupling potential energy factor increases, and increases when the cross-domain coupling potential energy factor decreases. The sensitivity configuration specifies the strength of the response of the temperature change amplitude to the change in potential energy factor. During the temperature determination process, the cross-domain coupling potential energy factor is first constrained to ensure that it falls within the effective numerical range. Then, the cross-domain coupling potential energy factor is mapped to temperature parameters according to the temperature adjustment rules, and boundary clipping is performed on the mapping results to ensure that the temperature parameters are always between the minimum and maximum temperature values. The original node quantity scores are scaled and probabilistically distributed using adaptive temperature parameters to obtain a node quantity probability distribution. The node quantity with the highest probability is selected as the optimal node quantity. Specifically, the original node quantity score sequence is scaled according to the adaptive temperature parameters. The scaling logic controls the amplification or compression of score differences based on the temperature parameters: smaller temperature parameters enhance score differences, while larger temperature parameters weaken score differences. Then, the scaled node quantity score sequence is probabilistically distributed using a normalized exponential allocation strategy, ensuring that each candidate node quantity category corresponds to a probability value and that the sum of all probability values ​​is one. After probabilistic distribution, a node quantity probability distribution is obtained. The candidate node quantity category with the highest probability value in the node quantity probability distribution is selected as the optimal node quantity, and the optimal node quantity and the node quantity probability distribution are written together into the output structure. When multiple candidate node quantity categories have the same probability value, all of which are the maximum value, a preset tie-breaking rule is used to determine the optimal node quantity. The tie-breaking rule can be configured to select the candidate category with the smaller node quantity or the category with the earlier candidate category order. Based on the adaptive temperature parameters, the original node adaptation scores are subjected to temperature scaling and probabilistic processing to obtain the node adaptation probability distribution. Specifically, the original node adaptation score sequence is scaled using a temperature scaling strategy consistent with the node quantity probability distribution to maintain temperature consistency in the node quantity decision and node selection decision for the same task. Subsequently, the scaled node adaptation score sequence is probabilistically processed using the same normalized exponential allocation strategy as described above to obtain the node adaptation probability distribution, where each probability value corresponds one-to-one with a node identifier in the candidate node set through a sequence index. Based on the optimal number of nodes, the node adaptation probability distribution is sorted by probability magnitude. Node identifiers that are ranked high and whose quantity equals the optimal number of nodes are selected. Specifically, the node adaptation probability distribution is sorted from highest to lowest probability value, maintaining the binding relationship between "probability value - node identifier" throughout the sorting process. The selection quantity is determined according to the optimal number of nodes, and node identifiers with high probability values ​​and sufficient quantity are selected sequentially from the sorted results to form a node identifier set.

[0030] The present invention is further configured such that obtaining the node allocation result includes: Node resource status data is obtained based on node identifiers. This data includes node type, current load, region, resource capacity, and operating status. Specifically, the node identifier is used as the query key to retrieve the corresponding node resource status data from the node resource status database. Node resource status data is indexed by node identifier, and each record contains at least node type, current load, region, resource capacity, and operating status. Node type is used to determine if the node matches the task's constraints on node categories; current load indicates the node's current resource occupancy level; region aligns with the task's operating region requirements; resource capacity verifies whether the node's available resources meet the task's needs; and operating status indicates whether the node is in a schedulable state. Based on the node resource status data, availability checks are performed on the nodes corresponding to the node identifiers, generating a set of unavailable node identifiers. Specifically, availability checks are performed on each node in the node identifier set sequentially. Availability checks employ a multi-condition rule set, which is fixed in the system configuration and associated with the task requirement field. This rule set includes at least four criteria: running status determination, region matching determination, capacity fulfillment determination, and load threshold determination. Running status determination requires the node's running status to be within the schedulable status set; region matching determination requires the node's region to meet the task's running region requirements; capacity fulfillment determination requires the node's resource capacity to meet the task's resource capacity requirements; and load threshold determination requires the node's current load to not exceed the preset load limit. After performing the above checks on each node, a verification result flag is generated, containing either a pass or fail status, and recording the rule item flag that triggered the failure. The node identifier corresponding to the fail status is added to the unavailable node identifier set, while the node identifier corresponding to the pass status is added to the temporarily stored available node identifier set. When the set of unavailable node identifiers is not empty, candidate replacement node identifiers are selected based on the node adaptation probability distribution, and node replacement is performed to obtain the set of replaced node identifiers. Specifically, when the set of unavailable node identifiers is not empty, the node adaptation probability distribution and the sequential snapshot identifiers of the candidate node set are read, maintaining the index binding relationship between the probability distribution and the node identifiers. First, node identifiers already in the temporarily stored set of available node identifiers are removed from the node adaptation probability distribution to avoid duplicate allocation; then, node identifiers in the set of unavailable node identifiers are removed to avoid reselecting failed nodes; then, pre-screening is performed on the remaining candidate node identifiers according to task requirements. The pre-screening conditions are consistent with or a subset of the availability verification rules in the second step to reduce the number of replacement iterations. The pre-screened candidate nodes are sorted from high to low according to their adaptation probability, and the node identifiers with the highest sorting value are selected as candidate replacement node identifiers according to the number of unavailable nodes. When performing a replacement, each candidate replacement node identifier is added to the set of available node identifiers in the temporary storage, and the availability check is performed again on the node after the addition, in the same way as in the second step; if the check fails, the replacement node is rolled back and the next candidate node is selected from the sorted list until the replacement is completed or the candidates are exhausted. The node allocation result is generated based on the replaced set of node identifiers and the optimal number of nodes. Specifically, the temporarily available node identifier set is sorted according to the node adaptation probability or the original order of candidate nodes, and the sorting rule is fixed in the system configuration. A corresponding number of node identifiers are extracted from the temporarily available node identifier set according to the optimal number of nodes to form the replaced node identifier set. The optimal number of nodes and the replaced node identifier set are then encapsulated into the node allocation result.

[0031] The present invention is further configured such that executing tasks according to node allocation results, monitoring task execution progress in real time, and collecting execution status feedback include: Based on the node allocation results, tasks are distributed to the corresponding nodes. Specifically, the node allocation results are read, which include the task identifier, the optimal number of nodes, and the set of allocated node identifiers. A distribution target list is established for each node in the allocated node identifier set, with each item in the target list including the node identifier, node communication address, scheduling channel identifier, and distribution sequence number. The distribution protocol type is determined based on the node type, which may include remote procedure call, message queue, or task queue methods. The task payload structure is determined based on the task type, and the task payload structure includes at least task input data, execution parameters, timeout control information, and result return address. The distribution operation is performed on each target node. Before distribution, a distribution signature identifier is generated and written to the distribution record, which includes the task identifier, node identifier, distribution time, payload summary, protocol type, and distribution status code. If distribution fails, a retry is performed according to a preset retry strategy, and the number of retryes and the failure reason identifier are recorded. If the retry still fails, the corresponding node identifier is added to the distribution failure set, and an exception handling flag is triggered. Task execution progress data is collected based on the execution process after task issuance. Specifically, after successful task issuance, an execution session record is established for each target node. The execution session record uses the task identifier and node identifier as a joint key and includes the session start time, session status, and progress collection cycle configuration. Progress collection adopts either an event reporting method or a polling query method: the event reporting method requires nodes to report progress events according to preset progress nodes during the execution phase; the polling query method requires the monitoring side to query the node for the current execution phase according to the progress collection cycle. Progress data includes at least a phase identifier, phase start time, phase completion time, and phase completion degree. The phase identifier is preset by the task execution process, and the phase completion degree is represented by a preset scale or discrete levels. Consistency checks are performed on the collected progress data, including timestamp increment checks, phase order checks, and duplicate reporting deduplication checks; progress data that passes the checks is written into the task progress data structure. Execution status data is collected based on task progress data. Specifically, after acquiring task progress data, a status collection strategy is triggered according to the progress stage. The status collection strategy is mapped from the stage identifier and fixed in the configuration. Execution status data includes at least node resource usage status, task running status, and abnormal status markers: node resource usage status includes at least one of the following indicators: computing resource usage, memory usage, storage usage, or network usage; task running status includes at least one of the following states: running, completed, failed, or timed out; abnormal status markers are used to identify the type of running abnormality, including insufficient resources, node unreachable, execution abnormality, or result feedback abnormality. Status collection sources can be node-side heartbeat reports, system monitoring interfaces, or task runtime log summaries. After collection, the status data undergoes format standardization and threshold determination. Threshold determination uses different threshold sets based on task type and node type, and the indicator items that trigger the threshold are written into the abnormal entries. The standardized execution status data is written into the execution status data structure and stored in association with the task identifier, node identifier, stage identifier, and collection time. Execution status data is written to node resource status data for feedback. Specifically, records in the execution status data structure are grouped by node identifier to form node-level feedback records. These node-level feedback records are written to the node resource status database. The written content includes at least the node's current load update value, running status update value, recent task execution result flag, and recent exception type flag. A versioned update mechanism is used during writing, retaining a snapshot before the update and recording the update reason identifier, which comes from the exception entry or task running status. After writing is complete, a feedback completion flag is generated and associated with the task identifier. The feedback completion flag is used by subsequent tasks to identify the latest status when reading node resource status data. When writing to the node resource status database fails, the feedback record is temporarily stored in a buffer queue according to a preset rollback strategy, and a rewrite process is initiated. Simultaneously, the write failure reason identifier and the number of rewrite attempts are recorded to ensure the continuous implementation of the feedback chain.

[0032] Example 2: Please see Figure 2 This exemplary dynamic node allocation system based on the GPT model includes: The task feature encoding module obtains node feature dimensions according to task requirements and encodes the feature values ​​of the node feature dimensions into token identifiers to generate a task identifier sequence. The time-series feature encoding module extracts time-series feature vectors based on task timestamp data and historical execution record data; The GPT processing module generates a fused sequence representation based on the task identifier sequence and temporal feature vector, generates fused features through dual attention branches and cross-branch interaction matrices, and calculates the original score of the number of nodes, the original score of node adaptation, and the cross-domain coupling potential factor. The temperature scaling decision module generates the optimal number of nodes and node identifiers based on the original scores of the number of nodes, the original scores of node adaptation, and the cross-domain coupling potential energy factor. The node resource management module performs availability checks based on the optimal number of nodes and node identifiers, replaces unavailable nodes, and generates node allocation results. The results output and execution module executes tasks based on the node allocation results, monitors the task execution progress in real time, and feeds back the execution status to the node resource management module.

[0033] It should be noted that the dynamic node allocation system based on the GPT model provided in the above embodiments and the dynamic node allocation method based on the GPT model provided in the above embodiments belong to the same concept. The specific methods of execution of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the dynamic node allocation system based on the GPT model provided in the above embodiments can be configured to perform the above-mentioned functional allocation by different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0034] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic node allocation method based on the GPT model, characterized in that, include: Based on the task requirements, obtain the node feature dimensions, encode the feature values ​​of the node feature dimensions into token identifiers, and generate a task identifier sequence. Extract time-series feature vectors based on task timestamp data and historical execution record data; A fusion sequence representation is generated based on the task identifier sequence and temporal feature vector. Fusion features are generated through dual attention branches and cross-branch interaction matrices. The original score of the number of nodes, the original score of node adaptation, and the cross-domain coupling potential factor are calculated. Based on the original scores for the number of nodes, the original scores for node adaptation, and the cross-domain coupling potential factor, the optimal number of nodes and node identifiers are generated. Availability is verified based on the optimal number of nodes and node identifiers, and unavailable nodes are replaced to generate node allocation results. Execute tasks based on node allocation results, monitor task execution progress in real time, and collect execution status for feedback.

2. The dynamic node allocation method based on the GPT model according to claim 1, characterized in that, The generated task identifier sequence includes: Analyze the task requirements to obtain the necessary node characteristic dimensions, including node type requirements, load capacity requirements, operating area requirements, resource capacity requirements, and node status requirements. For each node's feature dimension, a set of unique feature values ​​of the same number is pre-defined, and an encoding mapping from feature values ​​to token identifiers is established; Based on the encoding mapping, the feature value corresponding to the feature dimension of each node is converted into a token identifier; The token identifiers are combined according to the preset arrangement order of the node feature dimensions to generate a task identifier sequence.

3. The dynamic node allocation method based on the GPT model according to claim 1, characterized in that, Extracting time-series feature vectors based on task timestamp data and historical execution record data includes: Task timestamp data includes the current task submission time and the execution time of the most recent similar task; Historical execution record data includes the number of times the same type of task was executed within a preset time window and the historical average execution time interval; Calculate the time interval feature based on the current task submission time and the execution time of the most recent similar task, and combine the time interval feature, the number of similar tasks executed within a preset time window and the historical average execution time interval to generate a time series feature vector; Normalization is performed on the temporal feature vector to obtain the three-dimensional temporal feature vector.

4. The dynamic node allocation method based on the GPT model according to claim 1, characterized in that, The generation of fused sequence representations based on task identifier sequences and temporal feature vectors includes: A task token embedding sequence is generated based on the task identifier sequence, and a time-series token embedding sequence is generated based on the time-series feature vector. A fusion sequence is generated by concatenating the task token embedding sequence and the time sequence token embedding sequence in a preset order; A position embedding is generated based on the token position of the fused sequence, and the position embedding is injected into the fused sequence to obtain the fused sequence representation.

5. The dynamic node allocation method based on the GPT model according to claim 5, characterized in that, The generation of fused features through dual attention branches and cross-branch interaction matrices includes: Based on the fused sequence representation, query vectors, key vectors, and value vectors are generated through shared projection; In the task feature branch, generate a task branch attention score matrix within the task token range based on the query vector and key vector; In the temporal feature branch, a temporal branch attention score matrix is ​​generated within the temporal token range based on the query vector and key vector. Fill the intersection area of ​​task tokens and time-series tokens according to the cross-branch interaction matrix, and generate a fusion attention matrix by combining the task branch attention score matrix and the time-series branch attention score matrix; The fused features are obtained by performing weighted convergence on the value vectors based on the fusion attention matrix.

6. The dynamic node allocation method based on the GPT model according to claim 5, characterized in that, The calculation of the original score for the number of nodes, the original score for node adaptation, and the cross-domain coupling potential factor includes: Based on the fusion features, the original score for the number of nodes is generated through the node number output header, and the original score for node adaptation is generated through the node adaptation output header. Generate task domain energy terms based on task branch attention score matrix, generate time domain energy terms based on time branch attention score matrix, and generate cross-domain energy terms based on cross-branch interaction matrix; The cross-domain coupling potential energy factor is generated based on the task domain energy term, the time-series domain energy term, and the cross-domain energy term.

7. The dynamic node allocation method based on the GPT model according to claim 6, characterized in that, Generating the optimal number of nodes and node identifiers includes: Calculate the adaptive temperature parameters based on the cross-domain coupling potential energy factor; Based on the adaptive temperature parameters, the original fraction of the number of nodes is subjected to temperature scaling and probabilistic transformation to obtain the probability distribution of the number of nodes, and the number of nodes with the highest probability is selected as the optimal number of nodes. The node adaptation probability distribution is obtained by performing temperature scaling and probabilization on the original node adaptation scores based on the adaptive temperature parameters. Based on the optimal number of nodes, the node adaptation probability distribution is sorted by probability magnitude, and the nodes that are ranked first and whose number is equal to the optimal number of nodes are selected.

8. The dynamic node allocation method based on the GPT model according to claim 1, characterized in that, The generated node allocation results include: Node resource status data is obtained based on the node identifier. The node resource status data includes node type, current load, region, resource capacity, and operating status. Based on the node resource status data, perform availability checks on the nodes corresponding to the node identifiers and generate a set of unavailable node identifiers. When the set of unavailable node identifiers is not empty, candidate replacement node identifiers are selected according to the node adaptation probability distribution, and node replacement is performed to obtain the set of replaced node identifiers; The node allocation result is generated based on the replaced set of node identifiers and the optimal number of nodes.

9. A dynamic node allocation method based on the GPT model according to claim 8, characterized in that, Execute tasks based on node allocation results, monitor task execution progress in real time, and collect execution status feedback, including: Based on the node allocation results, the task will be distributed to the node corresponding to the node allocation result. Collect task execution progress data based on the execution process after the task is issued; Collect execution status data based on task execution progress data; The execution status data is written to the node resource status data for feedback.

10. A dynamic node allocation system based on the GPT model, used to implement the dynamic node allocation method based on the GPT model as described in any one of claims 1-9, characterized in that, include: The task feature encoding module obtains node feature dimensions according to task requirements and encodes the feature values ​​of the node feature dimensions into token identifiers to generate a task identifier sequence. The time-series feature encoding module extracts time-series feature vectors based on task timestamp data and historical execution record data; The GPT processing module generates a fused sequence representation based on the task identifier sequence and temporal feature vector, generates fused features through dual attention branches and cross-branch interaction matrices, and calculates the original score of the number of nodes, the original score of node adaptation, and the cross-domain coupling potential factor. The temperature scaling decision module generates the optimal number of nodes and node identifiers based on the original scores of the number of nodes, the original scores of node adaptation, and the cross-domain coupling potential energy factor. The node resource management module performs availability checks based on the optimal number of nodes and node identifiers, replaces unavailable nodes, and generates node allocation results. The results output and execution module executes tasks based on the node allocation results, monitors the task execution progress in real time, and feeds back the execution status to the node resource management module.

Images