Algorithm power matching method and system for large model cross-modal retrieval

Abstract

This invention provides a computing power matching method and system for cross-modal retrieval of large models, relating to the field of large model technology applications. First, it performs scenario-based computing power requirement mapping processing on the cross-modal retrieval requirements of large models, obtaining scenario-based computing power requirement results including basic computing power requirements and dynamic computing power requirements. Next, it constructs a resource supply capacity profile for each computing power node within a preset computing power pool. Then, based on the above results, it performs a dual-loop adaptation and adjustment process, obtaining an inner-loop local adaptation conclusion and an outer-loop global adaptation conclusion. Based on the dual-loop adaptation results, it generates a computing power adaptation judgment conclusion to indicate the adaptation status between scenario-based computing power requirements and resource supply capacity. Finally, based on the judgment conclusion, it generates computing power scheduling instructions to accurately allocate computing power tasks to the adapted computing power nodes, effectively improving the efficiency and performance of cross-modal retrieval of large models and reducing computing power costs.

Description

Technical Field

[0001] This invention relates to the field of large model technology applications, and more specifically, to a computational matching method and system for cross-modal retrieval of large models. Background Technology

[0002] With the rapid development of large-scale model technology, cross-modal retrieval, as a key application, has demonstrated immense value in numerous fields. For example, it enables the retrieval of image and text information in intelligent security, and assists doctors in quickly finding medical records related to specific images in the medical field. However, large-scale cross-modal retrieval places extremely high demands on computing power, and these demands are characterized by complexity and dynamism.

[0003] Existing computing power allocation methods are mostly simplistic and crude. On the one hand, they often fail to accurately assess computing power requirements based on different retrieval scenarios, and cannot distinguish between differences in basic computing power under different scenarios or dynamic computing power changes caused by fluctuations in retrieval load. This leads to unreasonable computing power allocation, either wasting computing power resources or failing to meet actual retrieval needs. On the other hand, they lack a comprehensive and detailed characterization of the resource supply capacity of each computing power node within the computing power pool, and cannot accurately grasp the current resource scale that a node can provide and the characteristics of continuous resource supply, thus failing to achieve efficient matching between computing power demand and supply. Furthermore, existing methods lack systematicity and flexibility in computing power adaptation and adjustment, making it difficult to dynamically adjust according to actual conditions. They cannot adapt to the ever-changing computing power requirements of large-scale cross-modal retrieval, seriously affecting retrieval efficiency and performance. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a computational matching method for cross-modal retrieval of large models, the method comprising: Scenario-based computing power requirement mapping is performed on the cross-modal retrieval requirements of large models to obtain scenario-based computing power requirement results. The scenario-based computing power requirement results include the basic computing power requirement corresponding to the retrieval scenario type and the dynamic computing power requirement corresponding to the retrieval load fluctuation. Construct a resource supply capacity profile for each computing node within a preset computing power pool. The resource supply capacity profile includes the current available resource scale and continuous resource supply characteristics of the computing power node. Based on the scenario-based computing power demand results and the resource supply capacity profile, a dual-loop adaptation and adjustment process is performed to obtain dual-loop adaptation results, which include the inner loop local adaptation conclusion and the outer loop global adaptation conclusion. A computing power adaptation judgment conclusion is generated based on the dual-loop adaptation results. The computing power adaptation judgment conclusion is used to indicate the adaptation status between scenario-based computing power demand and resource supply capacity. Based on the computing power adaptation determination conclusion, a computing power scheduling instruction is generated. The computing power scheduling instruction is used to allocate the computing power tasks corresponding to the cross-modal retrieval requirements of large models to the adapted computing power nodes.

[0005] In another aspect, embodiments of the present invention also provide a computing power matching system for cross-modal retrieval of large models, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the machine-readable storage medium to implement the above-described method.

[0006] Based on the above, this embodiment of the invention, by performing scenario-based computing power demand mapping processing, can comprehensively and accurately analyze the computing power demand under different retrieval scenarios, clearly distinguish and quantify the basic computing power demand and the dynamic computing power demand brought about by retrieval load fluctuations, avoiding resource waste or insufficiency caused by inaccurate computing power demand assessment. It constructs a resource supply capacity profile of each computing power node within a preset computing power pool, and performs a dual-loop adaptation and adjustment process based on the scenario-based computing power demand results and resource supply capacity profile. The inner loop local adaptation conclusion can quickly respond to local computing power changes and adjust local computing power allocation in a timely manner; the outer loop global adaptation conclusion grasps the overall supply and demand balance of computing power, ensuring that the resources of the entire computing power pool are optimally configured. A computing power adaptation judgment conclusion is generated based on the dual-loop adaptation results, and finally, a computing power scheduling instruction is generated based on the computing power adaptation judgment conclusion. This can accurately and efficiently allocate computing power tasks corresponding to large-model cross-modal retrieval needs to the adapted computing power nodes, significantly improving the efficiency and performance of large-model cross-modal retrieval, reducing computing power costs, and enhancing the stability and reliability of the system. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the computational matching method for cross-modal retrieval of large models provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the hardware architecture of a computing power matching system for cross-modal retrieval of large models provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a computational power matching method for cross-modal retrieval of large models provided in an embodiment of the present invention. The following is a detailed description of this computational power matching method for cross-modal retrieval of large models.

[0010] Step S110: Perform scenario-based computing power requirement mapping processing on the cross-modal retrieval requirements of large models to obtain scenario-based computing power requirement results. The scenario-based computing power requirement results include the basic computing power requirement corresponding to the retrieval scenario type and the dynamic computing power requirement corresponding to the retrieval load fluctuation.

[0011] In practice, when receiving a request for cross-modal retrieval of a large model, the first step is to comprehensively analyze the request to determine its corresponding scenario characteristics. This process requires extracting key information from the retrieval request, including the types of modalities involved in the retrieval, the target output format of the retrieval task, and the specific requirements for retrieval response speed.

[0012] After determining the retrieval scenario type, further analysis is conducted on the various operational steps required to complete the retrieval task within that scenario, clarifying the computational power consumption characteristics of each operational step. Simultaneously, historical retrieval load data for this scenario is collected, and the correlation between load changes and computational power consumption is analyzed to determine the baseline range of basic computational power requirements and the fluctuation range of dynamic computational power requirements.

[0013] Step S111: Analyze the retrieval scenario information of the cross-modal retrieval requirements of the large model, and determine the retrieval scenario type. The retrieval scenario type is divided according to the modal combination of the retrieval data, the output format of the retrieval results, and the timeliness requirements of the retrieval response.

[0014] When analyzing retrieval scenario information for cross-modal retrieval requirements of large models, it is necessary to extract the modal combinations involved one by one. Modal combinations may include various different forms of combination such as text and image, text and audio, and image and audio, each of which corresponds to different processing logic and computing power requirements.

[0015] Next, the output format of the search results needs to be defined. The output format may include a list display of search results, a relational graph display, a multimodal fusion display, etc. Different output formats have different requirements for data processing complexity and final computational power consumption.

[0016] Next, the timeliness requirements for the search response are determined. Timeliness requirements are typically defined by the maximum allowed time from the initiation of a search task to the return of results. Different timeliness requirements directly impact the real-time supply capacity of computing resources. Based on the above three aspects of information and according to pre-defined classification rules, the search requirements are identified as specific search scenario types.

[0017] Step S112: Extract the retrieval operation process corresponding to the retrieval scenario type. The retrieval operation process includes modal data retrieval action, modal feature association action, cross-modal result integration action, and retrieval result verification action.

[0018] After determining the search scenario type, a preset search operation process template is retrieved for that type. First, the modal data retrieval action is executed, which involves obtaining the corresponding modal data from the data storage node according to the search requirements. This action requires interaction with the data storage system and involves data transmission and retrieval operation logic.

[0019] After the modal data retrieval is completed, the modal feature association process begins. In this process, features are extracted from the data of different modalities, and then a feature matching algorithm is used to establish the association between the features of different modalities, thereby achieving preliminary cross-modal data association.

[0020] After completing the modal feature association, cross-modal result integration is performed. The intermediate search results obtained from associations across different modalities are merged and organized according to preset integration rules to form a unified search result set. Finally, a search result verification process is executed, using preset verification indicators to check the accuracy and completeness of the integrated search results, ensuring that the search results meet the basic requirements of the search needs.

[0021] Step S113: Query the preset scenario and computing power association library to obtain the standard computing power consumption benchmark corresponding to each search operation process under the search scenario type. The standard computing power consumption benchmark is the computing power consumption reference of the search operation process in a typical scenario.

[0022] The pre-defined scenario and computing power association database stores the correspondence between various search scenario types and the computing power consumption of their corresponding search operation processes. During the query process, the database is first indexed based on the determined search scenario type to locate the entry corresponding to that scenario type.

[0023] Under this entry, following the sequence of actions in the retrieval process, the standard computing power consumption benchmarks for each of the following actions are sequentially queried: modal data retrieval, modal feature association, cross-modal result integration, and retrieval result verification. Each standard computing power consumption benchmark includes information such as the computing power consumption range, main influencing factors, and reference calculation methods for that action under typical operating environments and data scales. The standard computing power consumption benchmarks for each queried action are recorded to ensure a one-to-one correspondence with the actions in the retrieval process.

[0024] Step S1131: Determine the storage structure of the preset scenario and computing power association library. The storage structure is classified according to the search scenario type. Under each search scenario type, there is a corresponding search operation process list and a standard computing power consumption benchmark for each process.

[0025] The pre-defined scenario and computing power association library adopts a hierarchical storage structure. The top layer is mainly classified according to the search scenario type, dividing different search scenario types into independent storage units. Under each search scenario type unit, there are sub-units for setting up a search operation process list and a standard computing power consumption benchmark.

[0026] The retrieval operation process list sub-unit stores the names of all retrieval operation actions under this scenario type and the logical relationships between the actions, arranged in the order of operation. The standard computing power consumption benchmark sub-unit corresponds to each action in the retrieval operation process list, storing the standard computing power consumption data for each action, including the calculation dimension, reference range, and influencing factors of the computing power consumption. This storage structure ensures that the required computing power consumption benchmark data can be quickly and accurately located during a query.

[0027] Step S1132: Based on the retrieved scenario type obtained from the parsing, locate the corresponding category entry in the scenario and computing power association library, and extract the retrieval operation process list under the category entry. The retrieval operation process list includes the names and order of all necessary retrieval operation processes under the scenario type.

[0028] Based on the previously parsed search scenario type name, a matching search is performed in the top-level category of the scenario and computing power association library. After finding the corresponding category entry, the search operation process list sub-unit under that category entry is opened.

[0029] Retrieve the stored retrieval operation process information from the retrieval operation process list sub-unit, extract all necessary retrieval operation process names, and record the sequence and dependencies between each operation process. For example, ensure that modal data retrieval actions precede modal feature association actions, and cross-modal result integration actions follow modal feature association actions. Organize the extracted operation process names and sequences into a structured list format.

[0030] Step S1133: For each retrieval operation process name, query the corresponding standard computing power consumption benchmark record under the category entry. The standard computing power consumption benchmark record includes the computing power consumption parameters, parameter units, and parameter determination basis of the retrieval operation process in typical scenarios.

[0031] For each operation process name in the compiled list of retrieval operation processes, a query is performed one by one in the standard computing power consumption benchmark sub-unit under the corresponding retrieval scenario type category. By matching the operation process name with the associated fields of the benchmark record, the corresponding standard computing power consumption benchmark record is found.

[0032] Each standard computing power consumption benchmark record includes various computing power consumption parameters for the operation process under typical scenarios, such as computing resource usage and computing power requirements corresponding to data processing throughput. The unit of each parameter is clearly recorded to ensure consistency in measurement. Furthermore, the basis for parameter determination is recorded, including data source, test environment, and sample size, to ensure the reliability and reference value of the standard computing power consumption benchmark.

[0033] Step S1134: Verify the completeness of the standard computing power consumption benchmark record. If the standard computing power consumption benchmark record is missing computing power consumption parameters or parameter units, retrieve the corresponding supplementary parameters from the historical supplementary records of the scenario and computing power association library.

[0034] After obtaining the standard computing power consumption baseline records for each operation process, the completeness of the records is checked. The check includes whether all necessary computing power consumption parameters are included, whether each parameter has a clear unit label, and whether the basis for parameter determination is complete.

[0035] When a missing computing power consumption parameter or an unclear unit is found in a standard computing power consumption benchmark record, a supplementation mechanism is activated. The system searches a specially stored historical supplementary record in the scenario-computing power association database, indexing the missing parameter by the operation process name and scenario type to retrieve the corresponding supplementary parameter. After obtaining the supplementary parameter, it is integrated into the original standard computing power consumption benchmark record, ensuring the completeness and validity of the computing power consumption benchmark record for each operation process.

[0036] Step S1135: Organize the standard computing power consumption benchmarks for each retrieval operation process according to the retrieval operation sequence to form a correspondence table between the retrieval operation process and the standard computing power consumption benchmarks under this retrieval scenario type.

[0037] After completing the standard computing power consumption benchmark query and integrity verification for all retrieval operation processes, each operation process name is associated and arranged with its corresponding standard computing power consumption benchmark according to the order determined in the retrieval operation process list.

[0038] The above correspondence is presented in tabular form. One column lists the names of the retrieval operation processes, and the other column lists the key parameters and ranges of the standard computing power consumption benchmark for each operation process. This arrangement creates a correspondence table, facilitating subsequent summarization, calculation, and retrieval of basic computing power requirements.

[0039] Step S114: Summarize the standard computing power consumption benchmarks of all search operation processes under the search scenario type to obtain the basic computing power requirements corresponding to the search scenario type.

[0040] After obtaining the correspondence table between the retrieval operation process and the standard computing power consumption benchmark, the standard computing power consumption benchmarks for all retrieval operation processes in the table are summarized. During the summary, the computing power consumption parameters for each operation process are categorized and accumulated. For example, the computing resource usage parameters for all operation processes are integrated, and the computing power requirement parameters related to data processing are integrated.

[0041] During the aggregation process, it is necessary to ensure that the units of similar parameters in different operation procedures are consistent. If there are unit differences, they should be uniformly converted before aggregation. After aggregation, a set of comprehensive computing power parameters is obtained. This set of computing power parameters constitutes the basic computing power requirement corresponding to this search scenario type, representing the basic computing power resource scale required to complete the search task in this scenario under typical conditions.

[0042] Step S115: Collect historical retrieval load data of cross-modal retrieval requirements of large models, analyze the load fluctuation pattern caused by changes in the amount of retrieval data and adjustments in retrieval accuracy in the historical retrieval load data, extract the range of computing power consumption changes corresponding to the load fluctuations, and use the range of computing power consumption changes as the dynamic computing power requirement corresponding to the retrieval load fluctuations.

[0043] Historical retrieval load data corresponding to the cross-modal retrieval requirements of this large model are collected from the log storage module of the retrieval system via a data interface. The collected data includes information such as the amount of retrieval data in different time periods, retrieval precision settings, and corresponding computing power consumption records.

[0044] The collected historical data was analyzed to identify sample data where load changes were caused by variations in the amount of retrieval data and adjustments to retrieval precision. For changes in the amount of retrieval data, the proportional relationship between the increase or decrease in data volume and the change in computing power consumption was analyzed; for changes in retrieval precision, the trend of computing power consumption when precision was increased or decreased was analyzed.

[0045] Statistical analysis of these sample data revealed patterns in load fluctuations, such as the increase in computing power consumption when the data volume increases by a certain percentage, and the range of computing power consumption changes when the precision is adjusted to different levels. Based on these patterns, the range of computing power consumption changes corresponding to load fluctuations was extracted and determined as the dynamic computing power requirement for this retrieval scenario.

[0046] Step S116: Integrate the basic computing power requirements corresponding to the search scenario type and the dynamic computing power requirements corresponding to the search load fluctuation to form a scenario-based computing power requirement result that includes the basic computing power requirements corresponding to the search scenario type and the dynamic computing power requirements corresponding to the search load fluctuation.

[0047] The previously obtained basic and dynamic computing power requirements are integrated and processed. First, the parameters of the basic computing power requirements are defined as baseline values. Then, the range of variation of the dynamic computing power requirements is correlated with the corresponding parameters of the basic computing power requirements.

[0048] For example, the computing resource usage parameter in the basic computing power requirement corresponds to a baseline range, while the dynamic computing power requirement specifies the range of increase or decrease of this parameter during load fluctuations. During integration, these relationships are presented in a structured data format, along with metadata such as the retrieval scenario type and data collection time range. The resulting scenario-based computing power requirement fully encompasses all aspects of both the basic and dynamic computing power requirements.

[0049] Step S120: Construct a resource supply capability profile for each computing node in the preset computing power pool. The resource supply capability profile includes the current available resource scale and continuous resource supply characteristics of the computing power node.

[0050] First, the scope of the preset computing power pool is defined, clearly identifying all computing power nodes included in the profile construction. Then, the current resource status of each computing power node is collected in real time through a resource monitoring interface, obtaining information on the current availability of computing resources, storage resources, network resources, etc.

[0051] Based on the current available resource scale, further collect historical resource supply data, resource replenishment mechanism information, and fault recovery records for each computing node. Analyze this data to evaluate the computing node's characteristics in terms of the timeliness and stability of resource replenishment, as well as its resource adjustment capabilities under load changes. Integrate the current available resource scale and the analyzed resource sustainability characteristics to create an independent resource supply capability profile for each computing node. The profile includes detailed information such as node identifier, resource type, current availability, resource replenishment method, and stability indicators.

[0052] Step S121: Determine the range of computing nodes in the preset computing power pool, and obtain the basic information of all computing nodes within the range of computing nodes. The basic information of the nodes includes the hardware configuration identifier, the resource cluster identifier, and the node running status identifier of the computing node.

[0053] The range of computing nodes in a preset computing power pool is determined by a preset configuration file, which records the IP addresses or unique device identifiers of all nodes belonging to that computing power pool. Based on this identifier information, the management interfaces connecting to each computing power node are traversed.

[0054] The management interface sends basic information query requests to each computing node to obtain the node's hardware configuration identifier, including identifier codes for hardware parameters such as processor model, memory capacity, and graphics card model; it also obtains the resource cluster identifier to determine which resource cluster the node belongs to for unified management; and it obtains the node's operating status identifier, such as normal operation, light load, heavy load, or fault status information. The obtained information is then categorized and organized by node to form a list of basic node information.

[0055] Step S122: Send a resource status collection request to each computing node. The resource status collection request includes the resource category to be collected, the data collection period, and the data feedback format.

[0056] Based on the need to construct a resource supply capacity profile, determine the resource categories that need to be collected, including specific categories such as computing resources, storage resources, and network resources. For each resource category, set a data collection cycle. For example, the collection cycle for computing resources can be set to a shorter interval to ensure real-time performance, while the collection cycle for storage resources can be appropriately extended.

[0057] Simultaneously, the format for data feedback is specified, including requirements for data field names, data types, and units, to ensure that the data format returned by each node is consistent. Based on the above settings, a resource status collection request is generated and sent to the resource monitoring module of each computing node via network communication protocol, awaiting the node to return resource status data.

[0058] Step S123: Receive resource status data returned by each computing node according to the resource status collection request. The resource status data includes the total resource capacity of each resource category, the amount of occupied resources, the trend of resource occupation, and information on resource replenishment channels.

[0059] Upon receiving a resource status collection request, the resource monitoring module of the computing node collects local resource status data according to the resource category and collection period specified in the request. After collection, the data is packaged according to the prescribed data feedback format and returned to the requesting party via network communication protocol.

[0060] After receiving the returned data, the receiver verifies the data's integrity and format correctness. If the verification passes, it parses the data content, extracting the total resource capacity for each resource category (i.e., the maximum available capacity of the node under that resource category); extracting the occupied resource quantity (i.e., the amount of resources currently in use); extracting the resource occupancy trend, typically represented by a resource occupancy rate change curve over a period of time; and extracting resource replenishment channel information, including the source, method, and triggering conditions of resource replenishment.

[0061] Step S124: Calculate the available resources for the resource status data. Subtract the occupied resources from the total resource capacity of each resource category to obtain the current available resources for each resource category. Summarize the current available resources for all resource categories to form the current available resource scale of the computing power node.

[0062] For each resource category, the total resource capacity data and the occupied resource amount data are calculated. Specifically, the occupied resource amount is subtracted from the total resource capacity to obtain the currently available resource amount for that resource category. For example, in a computing resource category, the total resource capacity is one value, and the occupied resource amount is another value; the difference between the two gives the currently available computing resource amount.

[0063] After performing the above calculation process on all resource categories, the current available resource quantities for each resource category are summarized. During the summarization, records are categorized according to resource category, forming structured data containing the current available quantities of computing resources, storage resources, network resources, etc. This structured data is the current available resource scale of the computing node.

[0064] Step S125: Analyze the resource replenishment channel information and resource occupation change trend in the resource status data to determine the resource replenishment method of the computing power node, the amount of resources that can be replenished within a unit period, and the interruption recovery characteristics of resource replenishment.

[0065] The information on resource replenishment channels is analyzed to clarify the specific ways in which computing power nodes obtain supplementary resources, such as through dynamic scheduling within the resource pool, external resource expansion, or the activation of redundant resources within the node itself.

[0066] By combining resource utilization trend data, we analyze the activation status of the resource replenishment mechanism when the resource utilization rate reaches different thresholds. Through analysis of historical replenishment records and current replenishment channel information, we determine the amount of resources that can be replenished within a unit period, that is, the upper limit of the amount of resources that a node can obtain through replenishment channels within a fixed time period.

[0067] At the same time, we analyze the possible interruptions that may occur during resource replenishment and the corresponding recovery measures, and determine the interruption recovery characteristics of resource replenishment, including the probability of interruption, the average recovery time after interruption, and the resource loss during the recovery process.

[0068] Step S126: Combining the historical resource supply records of computing power nodes, analyze the stability of resource replenishment methods, the consistency of the amount of resources that can be replenished within a unit period, and the reliability of interruption recovery characteristics, and integrate the analysis results to form the resource continuous supply characteristics of computing power nodes.

[0069] The historical resource supply records of the computing nodes are retrieved. These records contain detailed information such as resource replenishment operations, resource quantity changes, and interruption events over a past period. Based on these historical records, the stability of the resource replenishment methods is analyzed. Under the same conditions, the success rate and change frequency of each resource replenishment method are statistically analyzed. The lower the change frequency and the higher the success rate, the better the stability.

[0070] The consistency of available resources within a unit period is analyzed by comparing the deviation between the actual and expected replenishment amounts in different periods; a smaller deviation indicates higher consistency. Regarding the reliability of interruption recovery characteristics, indicators such as the percentage of successful recovery events and the stability of recovery time are statistically analyzed. Higher recovery success rates and more stable recovery times indicate better reliability. The results of these three analyses are integrated to form a comprehensive description of the continuous resource supply capability of the computing node, i.e., the continuous resource supply characteristic.

[0071] Step S1261: Retrieve the historical resource supply records of the computing power node. The historical resource supply records include records of changes in resource replenishment methods within a preset period, records of changes in the amount of replenishable resources within a unit period, and records of resource replenishment interruption and recovery times.

[0072] By accessing the historical data storage module of the computing node, the historical resource supply records of that node can be retrieved according to a preset time period. The preset period can be set according to actual needs, such as the past month or a quarter.

[0073] The historical resource supply records detail the changes in resource replenishment methods, including the time of each change, the previous method, the new method, and the reason for the change. The records of changes in the amount of replenishable resources within a unit period record the actual replenished resources, the planned replenished resources, and the difference between them for each period, in chronological order. The records of resource replenishment interruptions and recovery times detail the specific time of each interruption, the reason for the interruption, the duration of the interruption, the start time of recovery, the completion time of recovery, and the measures taken during recovery. These records are arranged chronologically to form a complete historical resource supply record dataset, providing raw data support for subsequent stability, consistency, and reliability analyses.

[0074] Step S1262: Analyze the resource replenishment method change records, and count the number of changes in the resource replenishment method within the preset period. The fewer the number of changes, the more stable the resource replenishment method meets the requirements.

[0075] The extracted resource replenishment method change records were reviewed one by one, and the total number of resource replenishment method changes within a preset time period was counted. The interval between each change was also recorded to analyze the frequency distribution of changes. If the number of changes within the preset period was low and the interval between changes was long, it indicated that the resource replenishment method maintained good stability during this period; if the number of changes was frequent and the interval between changes was short, it indicated that the resource replenishment method lacked stability. The counted number of changes was compared with a preset stability threshold. Changes below the threshold were considered to meet the stability requirements, while changes above the threshold were not, and specific instances of exceeding the threshold were recorded.

[0076] Step S1263: Analyze the change records of the amount of replenishable resources within a unit period, calculate the change range of the amount of replenishable resources within a preset period. The smaller the change range, the more consistent the amount of replenishable resources within a unit period is with the requirements.

[0077] For records of changes in replenishable resources within a unit period, the actual replenished resource data is extracted periodically. The deviation between the actual replenished resource amount in each period and the average actual replenished resource amount across all periods is calculated. Then, the overall change magnitude is calculated using these deviations. The change magnitude can be calculated by averaging the absolute values ​​of the deviations or by using the maximum deviation. The calculated change magnitude is compared with a preset consistency threshold. If the change magnitude is less than the threshold, the consistency of the replenishable resource amount within the unit period meets the requirements; the larger the change magnitude, the worse the consistency. For cases that do not meet the requirements, the specific period with the largest deviation and the corresponding deviation value must be recorded.

[0078] Step S1264: Analyze the time records of resource replenishment interruption and recovery, and count the total duration of resource replenishment interruption and the recovery duration after each interruption within the preset period. The shorter the total interruption duration and the shorter the recovery duration, the more the reliability of the interruption recovery characteristics meets the requirements.

[0079] Extract the duration and recovery time of each interruption from the resource replenishment interruption and recovery time records. Interruption duration is the interval between the interruption occurrence time and the recovery start time, and recovery time is the interval between the recovery start time and the recovery completion time. Sum the durations of all interruptions within a preset period to obtain the total interruption duration. Simultaneously, calculate the average and maximum recovery times after each interruption. A shorter total interruption duration indicates less impact on the resource replenishment process from interruptions; a shorter and less fluctuating recovery time indicates stronger recovery capability after an interruption. Compare the total interruption duration and average recovery time with corresponding reliability thresholds. If both are below the threshold, the reliability of the interruption recovery characteristics is deemed to meet the requirements; otherwise, the specific indicators and values ​​exceeding the thresholds must be recorded.

[0080] Step S1265: The stability analysis results of the resource replenishment method, the consistency analysis results of the replenishable resource amount within a unit period, and the reliability analysis results of the interruption recovery characteristics are correlated and integrated according to preset weights to form a comprehensive description.

[0081] Preset weights are assigned to the three analysis results: stability of resource replenishment methods, consistency of replenishable resource volume within a unit period, and reliability of interruption recovery characteristics. The weight values ​​are determined based on the importance of each characteristic in the continuous supply capacity of resources. The quantitative indicators of each analysis result (such as the number of changes, the magnitude of change, and the total interruption duration) are converted into corresponding score values. The range of score values ​​can be set to a fixed interval, for example, different scores corresponding to low to high.

[0082] The three scores are weighted according to preset weights to obtain a comprehensive score. Based on the comprehensive score and the preset scoring standards, a comprehensive description of the continuous resource supply capability of the computing node is generated. The comprehensive description must include key information from three aspects of the analysis results, such as stability score and number of changes, consistency score and magnitude of change, and reliability score and interruption recovery time, to comprehensively reflect the continuous resource supply characteristics.

[0083] Step S1266: Refer to the resource replenishment channel information in the current resource status data to correct the comprehensive description, and use the corrected comprehensive description as the resource continuous supply characteristic of the computing power node.

[0084] Retrieve resource replenishment channel information from the current resource status data and analyze whether the current replenishment channel is consistent with the historical replenishment channels. If discrepancies exist, the impact of these discrepancies on the continuous supply characteristics of resources needs to be assessed. For example, if the current replenishment channel is a newly activated channel, its stability may differ from historical channels, and this situation needs to be explained in the comprehensive description.

[0085] Simultaneously, the actual operating parameters of the current resource replenishment channels, such as the maximum replenishment capacity and response speed, are compared with historical analysis results. If there are significant deviations between the current parameters and historical data, the corresponding parts of the comprehensive description need to be corrected. For example, if the response speed of the current replenishment channel is better than that of historical channels, the relevant description of the reliability score can be appropriately increased. After the correction is completed, the final comprehensive description is the resource continuity supply characteristic of this computing power node, which is included in the resource supply capacity profile.

[0086] Step S127: Associate the current available resource scale and continuous resource supply characteristics of the computing power node with the node's basic information to construct a resource supply capacity profile that includes the current available resource scale and continuous resource supply characteristics of the computing power node.

[0087] The previously calculated available resource scale data is associated with the node identifier in the node's basic information to ensure accurate correspondence between the resource scale data and the node's data. Simultaneously, the analyzed resource sustainability characteristics are also associated with the corresponding node identifiers. During the association process, the integrity of the data and the accuracy of the correspondences must be checked to avoid data mismatches.

[0088] The basic information of a node, the current available resource scale, and the continuous resource supply characteristics are integrated into a structured dataset. This dataset is stored in a unified format and includes fields such as node hardware configuration identifier, resource cluster identifier, running status identifier, current available amount of various resources, stability description of resource replenishment methods, consistency description of replenishable resource quantity, and reliability description of interruption recovery. This structured dataset constitutes a resource supply capability profile for the computing power node. Once completed, it is stored in the computing power pool resource profile library for subsequent computing power matching processing.

[0089] Step S130: Based on the scenario-based computing power demand results and the resource supply capacity profile, perform a dual-loop adaptation adjustment process to obtain a dual-loop adaptation result, which includes the inner loop local adaptation conclusion and the outer loop global adaptation conclusion.

[0090] First, the internal loop local adaptation process is initiated. Based on the basic computing power requirements in the scenario-based computing power demand results, computing power nodes that can meet these basic requirements are selected from the resource supply capacity profile to form a candidate node group. The nodes in the candidate node group are further evaluated to determine whether their continuous resource supply characteristics meet the dynamic computing power requirements, generating an internal loop local adaptation conclusion.

[0091] If any unsuitable nodes exist within the inner loop, the outer loop global adaptation process is initiated. This expands the adaptation scope by adjusting the dynamic computing power demand description or supplementing the computing power node resource supply capacity profile. The inner loop adaptation steps are then executed again on the adjusted demands and supplemented profiles to select suitable node groups and analyze their resource distribution balance, generating the outer loop global adaptation conclusion. The adaptation conclusions from both the inner and outer loops are then integrated to form the final dual-loop adaptation result.

[0092] Step S131: Start the inner loop local adaptation process, extract the basic computing power requirement corresponding to a single search scenario type from the scenario-based computing power requirement results, and select computing power nodes that can currently provide resources to cover the basic computing power requirement from the resource supply capacity profile to form an inner loop candidate node group.

[0093] Extract the basic computing power requirement parameters corresponding to the single retrieval scenario type to be processed from the scenario-based computing power requirement results, including the baseline values ​​of the requirements for each resource category such as computing resources, storage resources, and network resources. Traverse the resource supply capacity profiles of all computing power nodes in the resource supply capacity profile library, and compare the current available resource scale and basic computing power requirement parameters of each node one by one.

[0094] During the comparison, for each resource category, it is determined whether the node's current available resource quantity is greater than or equal to the baseline value of the corresponding resource category in the basic computing power requirement. If the current available resource quantity for all resource categories can meet the basic requirements, the node is included in the inner loop candidate node group. If any resource category does not meet the requirements, the node is excluded. After the filtering is completed, the identifiers of all eligible nodes and their corresponding resource supply capacity profiles are summarized to form the inner loop candidate node group.

[0095] Step S132: For each computing power node in the inner loop candidate node group, based on its continuous resource supply characteristics, evaluate the range of computing power increment that the computing power node can provide in the future time period, and determine whether this range of computing power increment can meet the range of dynamic computing power demand changes corresponding to the search scenario type. If it can meet the requirements, mark it as an inner loop adaptable node; if it cannot meet the requirements, record the mismatched resource characteristic items.

[0096] For each computing node in the inner loop candidate node group, information such as the amount of resources that can be replenished within a unit period and the stability of resource replenishment are retrieved from its resource continuous supply characteristics. Combined with the preset time period length (such as one hour or one day in the future), the maximum and minimum computing power increments that can be provided within that time period are calculated based on the amount of resources that can be replenished within a unit period, forming a computing power increment range.

[0097] The calculated range of incremental computing power is compared with the range of dynamic computing power demand changes corresponding to the search scenario type in the scenario-based computing power demand results. The comparison includes whether the incremental demand of each resource category during dynamic changes is within the node's incremental computing power range. If the dynamic demand change range of all resource categories is covered by the node's incremental computing power range, the node is marked as an inner loop adaptation node. If the dynamic demand of some resource categories exceeds the node's incremental computing power range, the mismatched resource categories and their corresponding resource characteristics are recorded, such as insufficient replenishment per unit period or insufficient replenishment stability.

[0098] Step S133: Summarize the identification information of the adapted nodes in the inner loop, the matched resource characteristic items, and the mismatched resource characteristic items of the unadapted nodes to generate a partial adaptation conclusion for the inner loop. The partial adaptation conclusion for the inner loop includes a list of adapted nodes and the basis for adaptation.

[0099] Collect the identification information of all nodes marked as inner loop adaptable nodes, including the node's unique ID and the resource cluster to which it belongs. For each adaptable node, extract the resource characteristic items in its continuous resource supply characteristics that match the dynamic computing power demand, such as the unit period replenishment amount to meet dynamic incremental demand and the replenishment method that meets stability requirements, as the basis for adaptation.

[0100] Simultaneously, the identification information of unadapted nodes and their corresponding mismatched resource characteristics are summarized, detailing which resource categories each unadapted node is mismatched in, and the specific reasons for the mismatch, such as insufficient computing power increment range or substandard supplementary response speed. The list of adapted nodes, the adaptation basis, and the mismatch information of unadapted nodes are organized according to a preset format to form an inner-loop local adaptation conclusion. This inner-loop local adaptation conclusion will serve as the input basis for the outer-loop global adaptation processing.

[0101] Step S134: Initiate the outer loop global adaptation process. Based on the mismatched resource characteristics of the unadapted nodes in the inner loop local adaptation conclusion, adjust the dynamic computing power demand expression in the scenario-based computing power demand results, or coordinate the computing power nodes of other resource clusters in the preset computing power pool to supplement the resource supply capacity profile.

[0102] Analyze the resource mismatch characteristics of unfitted nodes in the inner loop local adaptation conclusions to summarize the main mismatch types and common problems. Based on the mismatch type, determine whether to adjust the dynamic computing power requirement statement or supplement computing power node resources. If the mismatch problem mainly stems from the dynamic computing power requirement statement being too strict or unreasonable, then initiate the dynamic computing power requirement statement adjustment process.

[0103] If the mismatch is due to insufficient node resources in the current resource supply capacity profile, a resource replenishment process is initiated. This process coordinates with other resource clusters within the preset computing power pool that are not included in the current profile to obtain their computing power node resource supply capacity information. This method expands the coverage of the resource supply capacity profile and increases the possibility of selecting suitable nodes.

[0104] Step S1341: Extract the mismatched resource characteristics of the unfitted nodes in the inner loop local adaptation conclusion, classify and statistically analyze the types of mismatched items, including insufficient continuous resource supply duration, insufficient dynamic replenishment amount, and insufficient replenishment response speed.

[0105] Extracting all mismatched resource characteristics from the internal loop local adaptation conclusions, and performing content analysis on each mismatch to determine its type. Insufficient continuous resource supply duration means the node cannot maintain a stable resource supply during the duration of dynamic demand; insufficient dynamic replenishment means the node's computing power increment is smaller than the range of dynamic demand changes; and inadequate replenishment response speed means that after dynamic demand is triggered, the node's resource replenishment initiation speed is slower than the response time specified by the demand.

[0106] Count the number of each type of mismatch and calculate its proportion in the total mismatches. For example, count how many nodes failed to adapt due to insufficient dynamic replenishment, and what percentage of this type it represents. By classifying and statistically analyzing these issues, we can identify the main types of mismatch problems, providing targeted directions for subsequent adjustments or supplementary measures.

[0107] Step S1342: If the mismatch type is insufficient dynamic replenishment or insufficient replenishment response speed, analyze the dynamic computing power demand expression in the scenario-based computing power demand results to determine if there are any abnormal expressions. If so, adjust the expression of the dynamic computing power demand range so that the adjusted expression is more in line with the current resource supply capacity.

[0108] For mismatches such as insufficient dynamic replenishment or inadequate replenishment response speed, re-examine the description of dynamic computing power requirements in the scenario-based computing power demand results. Check whether the range of changes in dynamic computing power requirements is based on reasonable historical data and actual needs, and whether there are any abnormalities in the description due to over-estimation or incorrect parameter settings. For example, is the required dynamic replenishment amount significantly higher than the actual maximum replenishment amount in the same period in history, or does the response speed requirement exceed the general level under current technical conditions?

[0109] If any discrepancies are found in the description, the upper limit of the dynamic computing power demand range or the response speed threshold will be adjusted based on the continuous resource supply characteristics of most nodes in the current resource supply capacity profile. The adjusted description should, while meeting the basic functions of retrieval, be as compatible as possible with the current resource supply capacity. For example, the upper limit of the dynamic replenishment demand may be appropriately reduced, or the allowable response time may be extended, so that more nodes can meet the adjusted dynamic computing power demand.

[0110] Step S1343: If the mismatch type is insufficient resource supply duration, query the computing node information of other resource clusters in the preset computing power pool and filter the computing nodes with longer resource supply duration.

[0111] When the main mismatch type is insufficient resource continuity duration, access the resource cluster management system of the preset computing power pool and query the list of computing power nodes of other resource clusters that are not currently included in the resource supply capacity profile. Obtain the resource continuity characteristics information of these nodes, focusing on the resource continuity duration indicator.

[0112] Based on the resource supply duration requirement standard in the internal loop adaptation, computing power nodes whose resource supply duration can meet or exceed the standard are selected. During the selection process, it is necessary to verify whether the current operating status of these nodes is normal and whether they meet the conditions for being added to the resource supply capacity profile. The node identifiers and their respective resource cluster information are recorded to prepare for subsequent resource status collection.

[0113] Step S1344: Send a resource status collection request to the computing power nodes of other selected resource clusters, obtain the resource status data of the nodes, and repeat the step of building a resource supply capacity profile to form a supplementary resource supply capacity profile.

[0114] Following the same resource status collection request format as before, send collection requests to the computing power nodes of other selected resource clusters, specifying the resource categories to be collected, the data collection period, and the data feedback format. Receive resource status data returned by these nodes, including total resource capacity, occupied resource amount, resource occupation trend, and resource replenishment information.

[0115] After verifying and parsing the received data, the process of constructing a resource supply capacity profile (steps S124 to S127) is repeated. This involves calculating the current available resource scale of these nodes, analyzing their continuous resource supply characteristics, and associating them with basic node information to ultimately form a supplementary resource supply capacity profile. The format of the supplementary resource supply capacity profile remains consistent with the original resource supply capacity profile to ensure seamless integration.

[0116] Step S1345: Integrate the supplementary resource supply capacity profile into the original resource supply capacity profile, and update the node coverage and resource characteristic description of the resource supply capacity profile so that the updated profile includes computing power nodes that can cover resource characteristic items that do not match.

[0117] Add each node from the supplementary resource supply capacity profile to the existing resource supply capacity profile database, ensuring the uniqueness of node identifiers and avoiding duplicate records. Update the node coverage statistics of the resource supply capacity profile, including the number of newly added nodes and their distribution across resource clusters.

[0118] For any inconsistencies between the resource characteristic descriptions in the original profile and the supplementary nodes, a unified description standard must be established to ensure that the resource characteristic descriptions for all nodes are presented in a consistent format and with uniform metrics. After the update, check whether the updated resource supply capacity profile includes computing power nodes that can cover previously mismatched resource characteristic items (such as insufficient resource continuous supply duration). If it does, the update is complete; otherwise, nodes need to be re-selected and supplemented.

[0119] Step S1346: Verify whether the adjusted scenario-based computing power demand result or the updated resource supply capacity profile can reduce the number of unfit nodes. If it can reduce the number, determine the adjustment or supplementation plan. If it cannot reduce the number, re-analyze the mismatched resource characteristics and adjust the processing method.

[0120] For the adjusted scenario-based computing power requirements, the filtering step in the inner loop local adaptation process is re-executed. The number of unadapted nodes is counted and compared with the number before the adjustment. If the number of unadapted nodes is significantly reduced, and the reduction ratio reaches a preset threshold, the adjustment scheme is considered effective, and the adjusted dynamic computing power requirement expression is adopted.

[0121] For the updated resource supply capacity profile, the internal loop local adaptation screening is re-executed to track changes in the number of mismatched nodes. If the number of mismatched nodes decreases due to the addition of new nodes, and the new nodes effectively cover the mismatched resource characteristics, then the supplementary solution is considered effective. If neither the adjustment nor the supplementary solution reduces the number of mismatched nodes, the underlying causes of the mismatched resource characteristics need to be re-analyzed. For example, are there any unconsidered resource characteristic influencing factors, or is it necessary to simultaneously take measures to adjust demand and supplement resources, and thus adjust the handling method accordingly.

[0122] Step S135: Repeat the inner loop local adaptation process for the adjusted scenario-based computing power demand results and the supplemented resource supply capacity profile, and select computing power nodes that can cover the adjusted basic computing power demand and dynamic computing power demand to form an outer loop adaptation node group.

[0123] The basic computing power requirement and dynamic computing power requirement in the adjusted scenario-based computing power requirement result are used as the new adaptation benchmark. The supplemented resource supply capacity profile is used as the source of node resource information. The internal loop local adaptation process of steps S131 to S133 is repeated.

[0124] First, nodes whose current resource availability can cover the adjusted basic computing power demand are selected to form a new candidate node group. Then, the resource supply characteristics of these candidate nodes are evaluated to determine if they meet the range of dynamic computing power demand changes after the adjustment. Nodes that meet the criteria are marked as outer loop adaptation nodes. The identification information and adaptation criteria of all outer loop adaptation nodes are collected to form the outer loop adaptation node group.

[0125] Step S136: Analyze the resource distribution balance of the outer loop adaptation node group within the preset computing power pool, and determine whether there is resource concentration or resource waste. If so, adjust the resource allocation ratio of the adaptation nodes; if not, determine the outer loop adaptation node group.

[0126] When analyzing the resource distribution of the outer loop adaptation node group within the preset computing power pool, it is necessary to consider multiple dimensions, including resource cluster affiliation, node geographical distribution (if distributed deployment is involved), and resource type proportion. First, the number of nodes and corresponding resource supply scale of each resource cluster in the outer loop adaptation node group are counted, and the resource proportion of each resource cluster in the entire node group is calculated. If the resource proportion of a certain resource cluster exceeds a preset threshold, such as exceeding half of the total resources, it is determined that there is resource concentration.

[0127] Simultaneously, the expected resource utilization rate of each adapted node is analyzed. Resource utilization is calculated by comparing the current available resource scale of a node with the amount of resources allocated to the retrieval task. If the resource utilization rate of multiple nodes is lower than a preset lower limit, such as below 30%, it is determined that there is resource waste.

[0128] If resource concentration or waste is identified, a resource allocation adjustment mechanism is activated. For resource concentration, the number of nodes allocated to clusters with excessively high resource usage is reduced, and some tasks are distributed to suitable nodes in other resource clusters. For resource waste, task allocations on some underutilized nodes are merged, or the resource allocation of nodes is adjusted to a reasonable range. After adjustment, the resource distribution is recalculated until the resource distribution balance requirement is met. If no resource concentration or waste exists, the outer loop suitable node group is directly determined as the final outer loop suitable node group.

[0129] Step S137: Summarize the identification information, resource allocation ratio and resource distribution balance analysis results of the outer loop adaptation node group, and generate the outer loop global adaptation conclusion.

[0130] Collect unique identifiers for all nodes in the outer loop adaptation node group, including node ID and the identifier of the resource cluster to which they belong. Compile resource allocation ratio data for each node, that is, the proportion of computing power tasks undertaken by the node in the entire outer loop adaptation node group, as well as the specific allocation ratio of various resources (computing resources, storage resources, and network resources).

[0131] The various indicators, judgment results, and adjustment records (if any) obtained during the resource distribution balance analysis are summarized, including the resource proportion of each resource cluster, node resource utilization statistics, and changes in resource allocation before and after the adjustment. Node identification information, resource allocation ratios, and resource distribution balance analysis results are integrated according to a preset format to form an outer-loop global adaptation conclusion. This outer-loop global adaptation conclusion fully reflects the adaptation status and resource allocation status of computing power nodes globally.

[0132] Step S138: Integrate the local adaptation conclusions of the inner loop and the global adaptation conclusions of the outer loop to form a double loop adaptation result that includes the local adaptation conclusions of the inner loop and the global adaptation conclusions of the outer loop.

[0133] The list of adapted nodes, the basis for adaptation, and the mismatched resource characteristics of unadapted nodes in the local adaptation conclusion of the inner loop are linked and integrated with the identification information of the adapted node group, the resource allocation ratio, and the resource distribution balance analysis results in the global adaptation conclusion of the outer loop.

[0134] During the integration process, a correspondence was established between the inner loop adaptation nodes and the outer loop adaptation nodes, clarifying which nodes were selected in both adaptation processes and which nodes were selected only in one adaptation. Simultaneously, the reasons for mismatches in the inner loop were correlated with the adjustment measures in the outer loop, demonstrating how adjustments in the outer loop resolved mismatches in the inner loop. The final dual-loop adaptation result contains complete information on both local and global adaptation, providing a comprehensive basis for subsequent computing power adaptation determination.

[0135] Step S140: Generate a computing power adaptation judgment conclusion based on the dual-loop adaptation result. The computing power adaptation judgment conclusion is used to indicate the adaptation status of scenario-based computing power demand and resource supply capacity.

[0136] The core indicators for local adaptation conclusions in the inner loop and global adaptation conclusions in the outer loop are extracted from the dual-loop adaptation results. The basic adaptation status of computing power requirements in a single scenario is determined based on the local adaptation conclusions in the inner loop, while the overall adaptation status of collaborative computing power requirements across multiple scenarios is determined based on the global adaptation conclusions in the outer loop. By combining the results of these two assessments, the adaptation status of scenario-based computing power requirements and resource supply capabilities is determined, including full adaptation, partial adaptation, or no adaptation. Corresponding node information, adaptation parameters, and unmet requirements are recorded to form a computing power adaptation judgment conclusion.

[0137] Step S141: Extract the inner loop local adaptation conclusion from the dual loop adaptation results, analyze the number of nodes in the inner loop adaptation node list and the completeness of the adaptation basis of each adaptation node, and determine whether the inner loop local adaptation meets the basic adaptation requirements of computing power demand in a single scenario.

[0138] From the dual-loop adaptation results, the inner loop local adaptation conclusion is extracted, with a focus on examining the list of inner loop adaptation nodes. The number of nodes in the list is counted; if the number of nodes is zero, it is directly determined that the inner loop local adaptation does not meet the basic adaptation requirements. If nodes exist, the adaptation basis for each node is checked one by one to ensure it is complete, including whether the basic computing power requirement coverage proof and dynamic computing power requirement satisfaction analysis are complete.

[0139] Set basic adaptation requirements and criteria, such as the number of nodes not less than a preset threshold and the completeness of adaptation criteria for all nodes reaching a preset percentage or higher. Compare the statistically measured number of nodes and the completeness of adaptation criteria with the criteria. If both are met, the inner loop local adaptation is deemed to meet the basic adaptation requirements for the computing power needs of a single scenario; if one or more of these are not met, the basic adaptation requirements are deemed not met.

[0140] Step S142: Extract the outer loop global adaptation conclusion from the dual loop adaptation results, analyze the rationality of the resource allocation ratio and the balance of resource distribution of the outer loop adaptation node group, and determine whether the outer loop global adaptation meets the overall adaptation requirements of multi-scenario collaborative computing power needs.

[0141] Extracting the global adaptation conclusions of the outer loop from the dual-loop adaptation results, the analysis focuses on the resource allocation ratio and resource distribution balance of the outer loop adaptation node group. It analyzes whether the resource allocation ratio conforms to the preset reasonable range, such as whether the resource allocation ratio of each node matches the node's resource supply capacity, and whether there are any abnormal nodes with excessively high or low allocation ratios.

[0142] Simultaneously, review the resource distribution balance analysis results to confirm whether any resource concentration or waste has been effectively addressed, whether the resource proportion of each resource cluster is within a reasonable range, and whether the node resource utilization rate meets the expected standards. Set judgment conditions for overall adaptation requirements: if the resource allocation ratio is reasonable and the resource distribution is balanced, then the outer loop global adaptation is determined to meet the overall adaptation requirements for multi-scenario collaborative computing power needs; otherwise, it is determined that the overall adaptation requirements are not met.

[0143] Step S143: If the local adaptation of the inner loop meets the basic adaptation requirements and the global adaptation of the outer loop meets the overall adaptation requirements, then it is determined that the scenario-based computing power demand and resource supply capacity are in a fully adapted state, and the node identifier and adaptation parameters of the fully adapted node are recorded.

[0144] When the judgment result of step S141 is that the local adaptation of the inner loop meets the basic adaptation requirements, and the judgment result of step S142 is that the global adaptation of the outer loop meets the overall adaptation requirements, the judgment of the fully adapted state is triggered.

[0145] In this state, the identification information of all nodes in the outer loop adaptation node group is collected to ensure that the identification of each node is accurate. Adaptation parameters corresponding to each node are extracted, including detailed parameters such as basic resource allocation, dynamic resource adjustment threshold, and resource allocation ratio. The above node identifications and adaptation parameters are recorded in a standardized format as proof of full adaptation and incorporated into the computing power adaptation judgment.

[0146] Step S144: If the local adaptation of the inner loop meets the basic adaptation requirements but the global adaptation of the outer loop does not meet the overall adaptation requirements, then it is determined that the scenario-based computing power demand and resource supply capacity are in a local adaptation state, and the node identifier, adaptation parameters and unmet global requirements of the local adaptation are recorded.

[0147] When the inner loop's local adaptation meets the basic adaptation requirements, but the outer loop's global adaptation does not meet the overall adaptation requirements, it is determined to be a local adaptation state. Nodes that meet the basic adaptation requirements are selected from the outer loop's adaptation node group, and their identification information is recorded.

[0148] Extract the adaptation parameters of these locally adapted nodes, including basic resource allocation and dynamic resource adjustment rules. Simultaneously, record in detail the overall requirements that were not met in the outer loop global adaptation, such as specific values ​​of resource allocation ratios exceeding reasonable ranges, resource cluster identifiers in resource concentrations, and lists of nodes with wasted resources. The node identifiers, adaptation parameters, and unmet global requirements of locally adapted nodes are then compiled and incorporated into the computing power adaptation judgment.

[0149] Step S1441: Confirm the criteria for determining whether the local adaptation of the inner loop meets the basic adaptation requirements. Specifically, the nodes in the list of adaptation nodes of the inner loop can cover the basic computing power requirements and dynamic computing power requirements of a single scenario, and the adaptation criteria are complete.

[0150] Review the analysis process of the conclusion of the local adaptation of the inner loop, and verify whether each node in the list of adaptation nodes of the inner loop can indeed cover the basic computing power requirements of a single scenario, that is, the parameters of the node's current available resource scale being greater than or equal to the basic computing power requirements.

[0151] Check whether the continuous resource supply characteristics of each node meet the changing range of dynamic computing power demand, ensuring that the nodes can provide sufficient computing power support during load fluctuations. Simultaneously, confirm that the adaptation basis for each node is complete, including whether the documentation such as basic computing power coverage proof, dynamic computing power fulfillment analysis, and resource characteristic matching descriptions is complete. Only when all nodes meet the above conditions can it be confirmed that the inner loop local adaptation meets the basic adaptation requirements.

[0152] Step S1442: Analyze the specific manifestations of the outer loop global adaptation not meeting the overall adaptation requirements. The specific manifestations include the resource allocation ratio exceeding the preset reasonable range, the resource distribution being concentrated in a certain proportion of resource clusters, or the computing power requirements of some scenarios not finding an adapted node.

[0153] A thorough analysis of the overall adaptation results for the external loop is conducted to identify the specific reasons why the overall adaptation requirements are not met. If the resource allocation ratio exceeds the preset reasonable range, it is necessary to identify which nodes have excessively high or low allocation ratios, and the specific extent of the excess.

[0154] If resources are concentrated in a specific area, identify the cluster identifier and its proportion in the total resources, and compare it with the preset distribution requirements. If no suitable nodes can be found for the computing power requirements of certain scenarios, record the type of these scenarios and the corresponding computing power requirement parameters, and analyze the specific reasons for not finding suitable nodes, such as insufficient resource supply or mismatched resource characteristics. Record the above specific manifestations in detail as the basis for not meeting the global requirements.

[0155] Step S1443: Extract the node identifier of the local adaptation node. The node identifier is a node in the inner loop adaptation node list that can meet the basic adaptation requirements but has resource allocation or distribution problems in the global adaptation.

[0156] Nodes with resource allocation or distribution issues were selected from the list of inner loop adaptation nodes. Although these nodes can meet the basic computing power requirements and dynamic computing power requirements of a single scenario, their resource allocation ratio is unreasonable or has led to problems such as concentrated resource distribution.

[0157] Extract the unique identifiers of these nodes, including node ID and the resource cluster they belong to, to ensure the accuracy and uniqueness of the identifiers. List these node identifiers separately as a set of identifiers for locally adapted nodes.

[0158] Step S1444: Query the adaptation parameters corresponding to the local adaptation node. The adaptation parameters include the basic resource allocation amount, dynamic resource adjustment rules, and the adaptation scenario type.

[0159] For the locally adapted nodes extracted in step S1443, the corresponding adaptation parameters are queried from the local adaptation conclusions of the inner loop and the global adaptation conclusions of the outer loop. The basic resource allocation amount is the quantity of various resources allocated to the node to meet the basic computing power requirements; the dynamic resource adjustment rules include the triggering conditions for resource adjustment (such as the load reaching a certain threshold), the adjustment direction (increase or decrease), and the adjustment magnitude; the adapted scenario type is the identifier of the search scenario type that the node can support.

[0160] The above adaptation parameters are organized by node to ensure that the parameters of each node are complete and accurate, providing a reference for subsequent computing power scheduling.

[0161] Step S1445: Record the unmet global requirements. If it is a resource allocation ratio problem, record the specific ratio that exceeds the value and the reasonable ratio range. If it is a resource distribution problem, record the centralized resource cluster identifier and the distribution requirements. If it is a partial scenario that is not adapted, record the incompatible scenario type and the corresponding computing power requirement.

[0162] Based on the specific performance analysis obtained in step S1442, classify and record the unmet global requirements. For the resource allocation ratio issue, record in detail the actual allocation ratio of each node, the upper and lower limits of the reasonable ratio range, and the specific value exceeding the limit (the difference between the actual ratio and the upper or lower limit of the reasonable range).

[0163] For resource distribution issues, record the resource cluster identifier, the resource percentage of that cluster, and the preset distribution requirements (such as the maximum allowed percentage). For issues where certain scenarios are not adapted, record the name or identifier of the incompatible scenario type, as well as the basic and dynamic computing power requirements corresponding to that scenario type. Organize these records by issue type to clearly present the unmet global requirements.

[0164] Step S1446: Organize the node identifiers, adaptation parameters, and unmet global requirements of the local adaptation according to the preset format to form a complete record of the local adaptation status.

[0165] Following a pre-defined document format, the list of node identifiers for local adaptation, the adaptation parameters for each node, and the unmet global requirements are structured and organized. The node identifier list is presented in tabular form, with two columns: node ID and the cluster to which it belongs. Adaptation parameters are listed separately for each node, with each node's parameters forming a separate section. Unmet global requirements are categorized by problem type, with detailed information listed under each problem type.

[0166] By using the above methods, a complete record of the local adaptation state is formed, ensuring that the record content is clear, well-organized, and easy to analyze and process later.

[0167] Step S145: If the local adaptation of the inner loop does not meet the basic adaptation requirements, it is determined that the scenario-based computing power demand and resource supply capacity are in an unadapted state, and the unadapted scenario type, the unmet basic computing power demand item and the corresponding resource gap information are recorded.

[0168] When the judgment result of step S141 is that the local adaptation of the inner loop does not meet the basic adaptation requirements, it is judged as an unadapted state. The retrieval scenario type corresponding to this unadapted state is clearly recorded, that is, the scenario type to which the current large model cross-modal retrieval requirement for computing power matching belongs.

[0169] The analysis identifies unmet basic computing power requirements in the internal loop local adaptation conclusions, specifically pinpointing which basic computing power requirements are not met, such as insufficient computing resources or inadequate storage resources. For each unmet basic computing power requirement, resource gap information is calculated, i.e., the difference between the actual available resources and the required resources. The misfit scenario types, unmet basic computing power requirements, and corresponding resource gap information are then compiled and incorporated into the computing power adaptation judgment conclusions.

[0170] Step S146: Integrate the judgment results, corresponding node identifiers, adaptation parameters, unmet requirements and resource gap information to generate a computing power adaptation judgment conclusion that indicates the adaptation status of scenario-based computing power demand and resource supply capacity.

[0171] The judgment result (fully adapted, partially adapted, or not adapted) obtained in steps S143, S144, or S145 is used as the core content, and the corresponding node identification information, adaptation parameters, unmet global requirements, or resource gap information are integrated.

[0172] Following a pre-defined conclusion template, the above information is structured to ensure completeness and logical clarity. Conclusions indicating full adaptation highlight the adaptation node identifiers and parameters; conclusions indicating partial adaptation emphasize local adaptation node information, parameters, and unmet global requirements; conclusions indicating non-adaptation detail the types of incompatible scenarios, unmet basic computing power requirements, and resource gaps. The resulting computing power adaptation judgment accurately and comprehensively reflects the adaptation status between scenario-specific computing power needs and resource supply capabilities.

[0173] Step S150: Generate a computing power scheduling instruction based on the computing power adaptation determination conclusion. The computing power scheduling instruction is used to allocate the computing power tasks corresponding to the cross-modal retrieval requirements of large models to the adapted computing power nodes.

[0174] Based on the adaptation status and related information in the computing power adaptation assessment, the adaptation nodes that need to be allocated computing power tasks are determined. For each adaptation node, a specific resource allocation plan is formulated based on the scenario-specific computing power requirements and adaptation parameters, including the basic resource allocation amount and dynamic resource adjustment rules. The above allocation plan is transformed into a standardized computing power scheduling instruction format, clarifying the execution target, execution content, and execution requirements of the instructions to ensure that computing power tasks can be accurately allocated to the adaptation computing power nodes.

[0175] For example, step S151: extract the identifiers of the adaptable nodes corresponding to the fully adapted state or the partially adapted state from the computing power adaptation determination conclusion to form a target node set, wherein the target node set contains the identifier information of at least one adaptable node.

[0176] Check the adaptation status of the computing power adaptation judgment result. If it is a fully adapted state, extract the identification information of all nodes in the outer loop adaptation node group; if it is a partially adapted state, extract the identification information of the locally adapted nodes.

[0177] The extracted node identifiers are deduplicated and verified to ensure that each identifier is valid and belongs to a node within the preset computing power pool. The verified node identifiers are then compiled into a list to form a target node set, which clearly defines the target node range for computing power task allocation.

[0178] Step S152: For each adaptable node in the target node set, query the corresponding adaptation parameters in the computing power adaptation judgment conclusion, and determine the basic resource allocation amount and time period for each adaptable node. The basic resource allocation amount is determined according to the basic computing power requirement in the scenario-based computing power requirement result.

[0179] Iterate through each compatible node in the target node set and look up the corresponding compatibility parameters in the computing power compatibility determination conclusion based on the node identifier. Extract the basic resource allocation from the compatibility parameters. This basic resource allocation is determined based on the basic computing power requirements in the scenario-based computing power requirement results and in combination with the resource supply capacity of the node, ensuring that the basic computing power requirements of a single scenario can be met.

[0180] Simultaneously, the time period for basic resource allocation is determined, namely the start and end times of resource allocation. The time period setting must cover the estimated execution time of the retrieval task. The basic resource allocation amount and time period for each node are recorded as the basis for computing power scheduling.

[0181] Step S153: Based on the dynamic computing power requirements in the scenario-based computing power requirements results, formulate dynamic resource adjustment rules for each adaptation node. The dynamic resource adjustment rules include the triggering conditions for resource adjustment and the method for determining the adjusted resource amount.

[0182] Referring to the dynamic computing power demand variation range in the scenario-based computing power demand results, dynamic resource adjustment rules are formulated for each adapted node. The triggering conditions are set based on the change indicators of retrieval load, such as when the amount of retrieval data exceeds a certain threshold, the retrieval accuracy is adjusted to a certain level, or the node resource utilization rate reaches a preset ratio, resource adjustment is triggered.

[0183] The method for determining the adjusted resource quantity needs to clearly define the calculation method for the direction (increase or decrease) and magnitude of the resource quantity adjustment when the triggering conditions are met. For example, the resource quantity can be adjusted according to the proportion of load change, or the resource quantity can be determined according to a preset step-by-step adjustment scheme. The triggering conditions and the method for determining the adjusted resource quantity should be integrated into the dynamic resource adjustment rules for each node.

[0184] Step S154: Based on the identification information of the adapting node, the basic resource allocation amount and time period, and the dynamic resource adjustment rules, construct the core content of the computing power scheduling instruction. The core content includes instruction identifier, task identifier, node identifier, basic resource configuration and dynamic adjustment rules.

[0185] A unique instruction identifier is generated for each computing power scheduling instruction, used for instruction tracking and management. This identifier is then associated with a corresponding retrieval task identifier, linking the instruction to a specific large-scale model cross-modal retrieval task.

[0186] The core content explicitly lists the identification information of the compatible nodes to ensure that instructions are accurately delivered to the target nodes. The basic resource allocation and time periods for each node are organized into the basic resource configuration section, and the dynamic resource adjustment rules are fully incorporated into the dynamic adjustment rules section. The core content constructed in this way comprehensively includes the key information required for computing power scheduling.

[0187] The instruction identifier is generated using specific encoding rules and includes a timestamp, task category code, and random sequence to ensure its uniqueness among all scheduling instructions in the preset computing power pool. The task identifier is consistent with the original task number of the large model cross-modal retrieval requirement, and can be directly linked to metadata such as the specific content, initiation time, and priority of the retrieval task.

[0188] The node identifiers are listed sequentially according to their order of appearance in the target node set. Each node identifier is on a separate line and includes the corresponding resource cluster identifier, facilitating quick identification of the node range involved in the command by the resource cluster management module. The basic resource configuration is presented in tabular form. The table columns include node identifier, computing resource allocation, storage resource allocation, network resource allocation, allocation start time, and allocation end time. The units for each type of resource allocation are consistent with those in the scenario-based computing power requirement results, and the allocation time period is accurate to the minute.

[0189] The dynamic adjustment rules section describes each node separately, first listing the node identifier, then explaining the various triggering conditions and corresponding adjustment methods for that node. Triggering conditions must clearly specify the indicator name, threshold range, and comparison method, such as "retrieved data volume > preset data volume threshold" or "node resource utilization rate ≥ preset utilization rate threshold." Adjustment methods must specify the type of resource adjustment (computing resources, storage resources, or network resources), the direction of adjustment, and the calculation method for the magnitude for each triggering condition, such as "when the retrieved data volume > preset data volume threshold, the computing resource allocation increases, and the increase is (actual data volume - preset data volume threshold) / preset data volume threshold × basic computing resource allocation."

[0190] Step S155: Supplement the execution monitoring requirements of the computing power scheduling instructions. The execution monitoring requirements include the monitoring frequency of resource usage status, the judgment criteria for abnormal situations, and the abnormal handling process.

[0191] The monitoring frequency of resource usage status is set according to the resource type and the sensitivity of dynamic adjustment rules. The monitoring frequency for computing resources is higher than that for storage and network resources. For example, computing resources are monitored once at shorter intervals, while storage and network resources are monitored once at longer intervals. The monitoring frequency is clearly recorded in the form of time intervals to ensure that the monitoring module collects node resource usage data at fixed periods.

[0192] The criteria for determining abnormal situations cover three scenarios: resource usage exceeding limits, resource allocation failure, and node response delay. The criteria for determining resource usage exceeding limits is that the actual usage of a certain type of resource by a node exceeds the preset proportion of the allocated amount multiple times consecutively. The criteria for determining resource allocation failure is that the node does not respond with resource configuration success information within a preset time after the command is issued. The criteria for determining node response delay is that the response time of the node in processing retrieval tasks exceeds the preset multiple of the average response time in this scenario multiple times consecutively.

[0193] The anomaly handling process employs a tiered handling mechanism for different anomalies. In scenarios where resource usage exceeds limits, dynamic resource adjustment rules are triggered for automatic intervention. If limits are still exceeded after several attempts, an alarm is sent to the resource management center. In resource allocation failure scenarios, scheduling instructions are automatically reissued. After multiple consecutive failures, the system switches to a backup node and records the faulty node information. In node response delay scenarios, the task allocation weight of the affected node is first reduced. If the delay persists, some tasks are migrated to other suitable nodes. All anomaly handling processes must clearly define the executing entity, operational steps, and information feedback requirements.

[0194] Step S156: Integrate the execution monitoring requirements into the core content to form a computing power scheduling instruction for allocating computing power tasks corresponding to the cross-modal retrieval requirements of large models to suitable computing power nodes.

[0195] The monitoring requirements are added as a separate chapter after the core content of the computing power scheduling instructions, using the same structured format as the core content. This includes a monitoring frequency table, an anomaly judgment criteria list, and a link to an anomaly handling flowchart. The monitoring frequency table is associated with the node identifiers and resource types in the core content, ensuring that each node has corresponding monitoring frequency records for each type of resource.

[0196] The list of anomaly judgment criteria is categorized by anomaly type. Each criterion entry includes judgment indicators, threshold parameters, and a description of the data source, which is explicitly stated as the node's real-time monitoring interface or a historical database query path. The anomaly handling flowchart links to a pre-set flowchart template library; the corresponding visual processing flowchart can be accessed via the flowchart number.

[0197] The final computing power scheduling instruction is encapsulated in a standardized XML format, comprising four parts: instruction header, core content, execution monitoring requirements, and signature information. The instruction header records the instruction generation time, the generating entity, and the version number; the signature information is a digital signature from the resource management center, used to verify the integrity and legitimacy of the instruction. The encapsulated computing power scheduling instruction is sent to the instruction receiving module of each adaptor node through an encrypted transmission channel, ensuring that the instruction is not tampered with or leaked during transmission.

[0198] Figure 2 The following is a schematic diagram of the hardware structure of a computing power matching system 100 for cross-modal retrieval of large models, provided by an embodiment of the present invention, for implementing the above-described computing power matching method for cross-modal retrieval of large models. Figure 2 As shown, the computing power matching system 100 for cross-modal retrieval of large models may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0199] In the specific implementation process, one or more processors 110 execute computer-executable instructions stored in machine-readable storage medium 120, so that processor 110 can execute the computing power matching method for cross-modal retrieval of large models as described in the above method embodiment. Processor 110, machine-readable storage medium 120 and communication unit 140 are connected through bus 130. Processor 110 can be used to control the sending and receiving actions of communication unit 140.

[0200] The specific implementation process of processor 110 can be found in the various method embodiments executed by the computing power matching system 100 for cross-modal retrieval of large models, as described above. The implementation principle and technical effect are similar, and will not be repeated here.

[0201] Furthermore, embodiments of the present invention also provide a readable storage medium containing computer-executable instructions. When the processor executes the computer-executable instructions, the above-described computational matching method for cross-modal retrieval of large models is implemented.

[0202] It should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof.

Claims

1. A computational matching method for cross-modal retrieval of large models, characterized in that, The method includes: Scenario-based computing power requirement mapping is performed on the cross-modal retrieval requirements of large models to obtain scenario-based computing power requirement results. The scenario-based computing power requirement results include the basic computing power requirement corresponding to the retrieval scenario type and the dynamic computing power requirement corresponding to the retrieval load fluctuation. Construct a resource supply capacity profile for each computing node within a preset computing power pool. The resource supply capacity profile includes the current available resource scale and continuous resource supply characteristics of the computing power node. Based on the scenario-based computing power demand results and the resource supply capacity profile, a dual-loop adaptation and adjustment process is performed to obtain dual-loop adaptation results, which include the inner loop local adaptation conclusion and the outer loop global adaptation conclusion. A computing power adaptation judgment conclusion is generated based on the dual-loop adaptation results. The computing power adaptation judgment conclusion is used to indicate the adaptation status between scenario-based computing power demand and resource supply capacity. Based on the computing power adaptation determination conclusion, a computing power scheduling instruction is generated. The computing power scheduling instruction is used to allocate the computing power tasks corresponding to the cross-modal retrieval requirements of large models to the adapted computing power nodes. The process of performing a dual-loop adaptation and adjustment based on the scenario-based computing power demand result and the resource supply capacity profile to obtain the dual-loop adaptation result includes: Initiate the internal loop local adaptation process, extract the basic computing power requirements corresponding to a single search scenario type from the scenario-based computing power demand results, and select computing power nodes that can currently provide resources to cover the basic computing power requirements from the resource supply capacity profile to form an internal loop candidate node group. For each computing power node in the inner loop candidate node group, based on its continuous resource supply characteristics, the range of computing power increment that the computing power node can provide in the future time period is evaluated, and it is determined whether this range of computing power increment can meet the range of dynamic computing power demand changes corresponding to the search scenario type. If it can meet the requirements, it is marked as an inner loop adaptable node; if it cannot meet the requirements, the mismatched resource characteristic items are recorded. The identification information of the adapted nodes in the inner loop, the matching resource characteristics, and the mismatched resource characteristics of the unadapted nodes are summarized to generate a partial adaptation conclusion for the inner loop. The partial adaptation conclusion for the inner loop includes a list of adapted nodes and the basis for adaptation. Initiate global adaptation processing for the outer loop, and based on the mismatched resource characteristics of the unadapted nodes in the conclusion of local adaptation for the inner loop, adjust the dynamic computing power demand expression in the scenario-based computing power demand results, or coordinate the computing power nodes of other resource clusters in the preset computing power pool to supplement the resource supply capacity profile. The process of performing the inner loop local adaptation process is repeated on the adjusted scenario-based computing power demand results and the supplemented resource supply capacity profile. Computing power nodes that can cover the adjusted basic computing power demand and dynamic computing power demand are selected to form an outer loop adaptation node group. Analyze the resource distribution balance of the outer loop adaptation node group within the preset computing power pool to determine whether there is resource concentration or resource waste. If so, adjust the resource allocation ratio of the adaptation nodes; if not, determine the outer loop adaptation node group. Summarize the identification information, resource allocation ratio, and resource distribution balance analysis results of the outer loop adaptation node group to generate a global adaptation conclusion for the outer loop; By integrating the local adaptation conclusions of the inner loop and the global adaptation conclusions of the outer loop, a dual-loop adaptation result is formed, which includes both the local adaptation conclusions of the inner loop and the global adaptation conclusions of the outer loop.

2. The computational power matching method for cross-modal retrieval of large models according to claim 1, characterized in that, The process of performing scenario-based computing power requirement mapping on the cross-modal retrieval requirements of large models to obtain scenario-based computing power requirement results includes: Analyze the retrieval scenario information of cross-modal retrieval requirements of large models, and determine the retrieval scenario type. The retrieval scenario type is divided according to the modal combination of retrieval data, the output format of retrieval results, and the timeliness requirements of retrieval response. Extract the retrieval operation process corresponding to the retrieval scenario type. The retrieval operation process includes modal data retrieval action, modal feature association action, cross-modal result integration action, and retrieval result verification action. Query the preset scenario and computing power association library to obtain the standard computing power consumption benchmark corresponding to each search operation process under the search scenario type. The standard computing power consumption benchmark is the computing power consumption reference of the search operation process in a typical scenario. By summarizing the standard computing power consumption benchmarks of all search operation processes under the aforementioned search scenario type, the basic computing power requirements corresponding to this search scenario type are obtained. Collect historical retrieval load data of cross-modal retrieval requirements of large models, analyze the load fluctuation pattern caused by changes in the amount of retrieval data and adjustments in retrieval accuracy in the historical retrieval load data, extract the range of computing power consumption change corresponding to the load fluctuation, and take the range of computing power consumption change as the dynamic computing power requirement corresponding to the retrieval load fluctuation. By integrating the basic computing power requirements corresponding to the search scenario types and the dynamic computing power requirements corresponding to search load fluctuations, a scenario-based computing power requirement result is formed, which includes the basic computing power requirements corresponding to the search scenario types and the dynamic computing power requirements corresponding to search load fluctuations.

3. The computational power matching method for cross-modal retrieval of large models according to claim 2, characterized in that, The query uses a preset scenario and computing power association library to obtain the standard computing power consumption benchmark corresponding to each search operation process under the search scenario type, including: The storage structure of the preset scenario and computing power association library is determined. The storage structure is classified according to the retrieval scenario type. Under each retrieval scenario type, a corresponding retrieval operation process list and a standard computing power consumption benchmark for each process are associated. Based on the retrieval scenario type obtained from the analysis, the corresponding category entry is located in the scenario and computing power association library, and the retrieval operation process list under the category entry is extracted. The retrieval operation process list includes the names and order of all necessary retrieval operation processes under the scenario type. For each retrieval operation process name, the corresponding standard computing power consumption benchmark record is queried under the category entry. The standard computing power consumption benchmark record contains the computing power consumption parameters of the retrieval operation process in typical scenarios, the parameter units, and the basis for parameter determination. Verify the completeness of the standard computing power consumption benchmark record found. If the standard computing power consumption benchmark record is missing computing power consumption parameters or parameter units, retrieve the corresponding supplementary parameters from the historical supplementary records of the scenario and computing power association library. The standard computing power consumption benchmarks for each retrieval operation process are organized according to the retrieval operation sequence to form a correspondence table between the retrieval operation process and the standard computing power consumption benchmarks under this retrieval scenario type.

4. The computational power matching method for cross-modal retrieval of large models according to claim 1, characterized in that, The process of constructing a resource supply capacity profile for each computing node within a pre-defined computing power pool includes: Determine the range of computing power nodes in the preset computing power pool, and obtain the basic node information of all computing power nodes within the range of computing power nodes. The basic node information includes the hardware configuration identifier, the resource cluster identifier, and the node running status identifier of the computing power node. Send a resource status collection request to each computing node. The resource status collection request includes the resource type to be collected, the data collection period, and the data feedback format. Receive resource status data returned by each computing node according to the resource status collection request. The resource status data includes the total resource capacity, occupied resource amount, resource occupation trend and resource replenishment channel information for each resource category. The available resources are calculated based on the resource status data. The total resource capacity of each resource category is subtracted from the occupied resource amount to obtain the current available resource amount of each resource category. The current available resource amounts of all resource categories are summarized to form the current available resource scale of the computing power node. Analyze the resource replenishment channel information and resource occupancy change trend in the resource status data to determine the resource replenishment method of the computing power node, the amount of resources that can be replenished within a unit period, and the interruption recovery characteristics of resource replenishment; By combining the historical resource supply records of computing power nodes, we analyze the stability of resource replenishment methods, the consistency of the amount of resources that can be replenished within a unit period, and the reliability of interruption recovery characteristics. The integrated analysis results form the resource continuous supply characteristics of computing power nodes. By associating the current available resource scale and continuous resource supply characteristics of computing power nodes with the node's basic information, a resource supply capability profile containing the current available resource scale and continuous resource supply characteristics of computing power nodes is constructed.

5. The computational power matching method for cross-modal retrieval of large models according to claim 4, characterized in that, The analysis combines historical resource supply records of computing nodes to determine the stability of resource replenishment methods, the consistency of replenishable resource amounts within a unit period, and the reliability of interruption recovery characteristics. The integrated analysis results form the resource sustainability characteristics of computing nodes, including: Retrieve historical resource supply records of computing power nodes. These historical resource supply records include records of changes in resource replenishment methods within a preset period, records of changes in the amount of replenishable resources within a unit period, and records of resource replenishment interruption and recovery times. Analyze the resource replenishment method change records and count the number of times the resource replenishment method changes within a preset period. The fewer the number of changes, the more stable the resource replenishment method meets the requirements. Analyze the records of changes in the amount of replenishable resources within a unit period, calculate the magnitude of changes in the amount of replenishable resources within a preset period, and the smaller the magnitude of changes, the more consistent the amount of replenishable resources within a unit period is with the requirements. Analyze the time records of resource replenishment interruption and recovery, and statistically analyze the total duration of resource replenishment interruption and the recovery time after each interruption within the preset period. The shorter the total interruption duration and the shorter the recovery time, the more the reliability of the interruption recovery characteristics meets the requirements. The stability analysis results of resource replenishment methods, the consistency analysis results of replenishable resources within a unit period, and the reliability analysis results of interruption recovery characteristics are correlated and integrated according to preset weights to form a comprehensive description. The comprehensive description is revised based on the resource replenishment channel information in the current resource status data, and the revised comprehensive description is used as the resource continuous supply characteristic of the computing power node.

6. The computational power matching method for cross-modal retrieval of large models according to claim 1, characterized in that, The mismatched resource characteristic items of the unadapted nodes in the conclusion of the inner loop local adaptation are used to adjust the dynamic computing power demand expression in the scenario-based computing power demand results, or to coordinate the computing power nodes of other resource clusters in the preset computing power pool to supplement the resource supply capacity profile, including: Extract the mismatched resource characteristics of the unfitted nodes in the local adaptation conclusion of the inner loop, classify and statistically analyze the types of mismatched items, and the types include insufficient continuous supply duration of resources, insufficient dynamic replenishment amount, and insufficient replenishment response speed. If the mismatch type is insufficient dynamic replenishment or the replenishment response speed does not meet the requirements, analyze the dynamic computing power demand expression in the scenario-based computing power demand results to determine whether there are any abnormal expressions. If so, adjust the expression of the range of change of dynamic computing power demand so that the adjusted expression is more in line with the current resource supply capacity. If the mismatch type is insufficient resource supply duration, query the computing node information of other resource clusters in the preset computing power pool and filter the computing nodes with longer resource supply duration. Send resource status collection requests to the computing power nodes of other selected resource clusters, obtain the resource status data of the nodes, and repeat the steps of building a resource supply capacity profile to form a supplementary resource supply capacity profile. The supplementary resource supply capacity profile is integrated into the original resource supply capacity profile, and the node coverage and resource characteristic description of the resource supply capacity profile are updated so that the updated profile can include computing power nodes that do not match the resource characteristic items. Verify whether the adjusted scenario-based computing power demand results or the updated resource supply capacity profile can reduce the number of unfit nodes. If it can reduce the number, determine the adjustment or supplementation plan. If it cannot reduce the number, re-analyze the mismatched resource characteristics and adjust the processing method.

7. The computational power matching method for cross-modal retrieval of large models according to claim 1, characterized in that, The step of generating a computing power adaptation determination conclusion based on the dual-loop adaptation result includes: Extract the local adaptation conclusions of the inner loop from the dual-loop adaptation results, analyze the number of nodes in the list of inner loop adaptation nodes and the completeness of the adaptation basis of each adaptation node, and determine whether the local adaptation of the inner loop meets the basic adaptation requirements of computing power demand in a single scenario. Extract the global adaptation conclusion of the outer loop from the dual loop adaptation results, analyze the rationality of the resource allocation ratio and the balance of resource distribution of the outer loop adaptation node group, and determine whether the global adaptation of the outer loop meets the overall adaptation requirements of multi-scenario collaborative computing power needs. If the local adaptation of the inner loop meets the basic adaptation requirements and the global adaptation of the outer loop meets the overall adaptation requirements, then the scenario-based computing power requirements and resource supply capabilities are determined to be in a fully adapted state, and the node identifier and adaptation parameters of the fully adapted node are recorded. If the local adaptation of the inner loop meets the basic adaptation requirements but the global adaptation of the outer loop does not meet the overall adaptation requirements, then the scenario-based computing power demand and resource supply capacity are determined to be in a local adaptation state, and the node identifier, adaptation parameters and unmet global requirements of the local adaptation are recorded. If the local adaptation of the inner loop does not meet the basic adaptation requirements, it is determined that the scenario-based computing power demand and resource supply capacity are in an unadapted state, and the unadapted scenario type, unmet basic computing power demand item and corresponding resource gap information are recorded. By integrating the judgment results, corresponding node identifiers, adaptation parameters, unmet requirements, and resource gap information, a computing power adaptation judgment conclusion is generated to indicate the adaptation status of scenario-based computing power demand and resource supply capacity.

8. The computational power matching method for cross-modal retrieval of large models according to claim 7, characterized in that, If the local adaptation of the inner loop meets the basic adaptation requirements but the global adaptation of the outer loop does not meet the overall adaptation requirements, then the scenario-based computing power demand and resource supply capacity are determined to be in a state of local adaptation. The node identifier, adaptation parameters, and unmet global requirements are recorded, including: The criteria for confirming that the local adaptation of the inner loop meets the basic adaptation requirements are as follows: the nodes in the list of adaptation nodes of the inner loop can cover the basic computing power requirements and dynamic computing power requirements of a single scenario, and the adaptation basis is complete. The analysis reveals specific manifestations of the failure of the outer loop global adaptation to meet the overall adaptation requirements. These specific manifestations include resource allocation ratios exceeding a preset reasonable range, resource distribution concentrated in resource clusters of a certain proportion, or the computing power requirements of certain scenarios failing to find suitable nodes. Extract the node identifier of the local adaptation node. The node identifier is a node in the inner loop adaptation node list that can meet the basic adaptation requirements but has resource allocation or distribution problems in the global adaptation. Query the adaptation parameters corresponding to the local adaptation node. The adaptation parameters include the basic resource allocation, dynamic resource adjustment rules, and the type of adaptation scenario. Record any unmet global requirements. If it is a resource allocation ratio issue, record the specific ratio that exceeds the value and the reasonable ratio range. If it is a resource distribution issue, record the centralized resource cluster identifier and the distribution requirements. If it is a scenario that is not adapted, record the type of scenario that is not adapted and the corresponding computing power requirements. The node identifiers, adaptation parameters, and unmet global requirements of the local adaptation are organized according to a preset format to form a complete record of the local adaptation status.

9. A computing power matching system for cross-modal retrieval of large models, characterized in that, The computational power matching system for cross-modal retrieval of large models includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the memory to implement the computational power matching method for cross-modal retrieval of large models as described in any one of claims 1-8.

Images