811 results about "Multiple Models" patented technology

The invention provides a real-time intelligent agent dynamic scheduling method and system composed of multiple large models, and is applied to the technical field of data processing. The method comprises the following steps: performing dynamic scheduling expansion around a plurality of large-model real-time intelligent agents, firstly receiving task request data of a real-time service scene, performing grouping according to task types, priorities and resource requirements, pre-configuring a scheduling strategy, and generating a pre-scheduling queue and resource pre-planning information; and in combination with the target configuration information, scheduling task information after secondary optimization is obtained through multi-dimensional optimization such as task splitting and priority rearrangement. Dynamically adjusting and generating a target scheduling instruction based on real-time system resources and a model operation state, and cooperatively processing the target scheduling instruction with the large model capability adaptation parameters to form scheduling execution information; and finally, issuing to model nodes according to preset service quality requirements and resource constraints, generating information such as adaptive parameters and priority rules through feature extraction, association processing and the like, and supporting multi-model cooperation to efficiently complete real-time service tasks.

The invention discloses a system and method for estimating the charging capacity of an energy storage lithium battery, and particularly relates to the technical field of energy storage battery management, and the method comprises the following steps: collecting dynamic parameters in the charging process of the energy storage lithium battery, and forming a dynamic data set; calculating an environmental disturbance influence index and a charging response consistency factor based on time sequence analysis and feature decoupling; constructing a two-dimensional working condition mapping matrix, identifying a current working condition area and generating a scene label; determining parameter weights of the plurality of estimation models by adopting a probabilistic reasoning mode; according to the scene state, selecting a single model output result or fusing a plurality of model output results for estimation; according to the method, the environment disturbance influence index and the charging response consistency factor are constructed, so that the complex working condition is accurately identified; scene labels are automatically generated based on two-dimensional working condition mapping and a clustering algorithm, and a plurality of estimation models are dynamically selected or fused in combination with model confidence, so that the accuracy, robustness and intelligent level of estimation are improved.

A vehicle-mounted GNSS positioning method based on multi-motion model interaction includes: establishing a position-constant velocity (PCV) model and a position-constant steering angular velocity (PCSAV) model for two attitudes of a carrier (i.e., linear motion and turning motion) respectively to obtain a state estimation vector and a state transition matrix of the carrier of the PCV model and the PCSAV model at a previous moment, introducing an interacting multiple model (INM), establishing a heuristic position-velocity filtering (HPV)-IMM model based on the IMM model to achieve an information filtering interaction between the PCV model and the PCSAV model, and obtaining a state estimation vector and an error covariance matrix of the carrier at a current moment, so as to obtain a position and velocity of the carrier at the current moment. The present disclosure solves the problem of low accuracy of a traditional single kinematic model in multi-motion attitude vehicle positioning.

The invention discloses a Java-based large language model dynamic adaptation method and system, belongs to the technical field of artificial intelligence, and aims to solve the technical problems of overcoming the interface difference of multiple model interfaces, simplifying the development process and providing standardized and high-expansibility large model management. Comprising the following steps: providing a model registration and dynamic loading service, a protocol conversion and parameter standardization service and a load balancing and failover service; a streaming transmission protocol service, a function call dynamic injection service and a global error processing mechanism are provided; when a user provides a document analysis and partitioning service, a vectorization index construction service and an RAG enhanced generation service to ask questions, relevant document blocks in the Pinecone are retrieved through the RAG enhanced generation service to serve as contexts to be injected into cue words of the large language model, and answers generated by the large language model are returned.

The invention discloses a power grid maintenance plan reliability post-evaluation method based on Monte Carlo simulation and data driving, and belongs to the technical field of power system operation and reliability analysis. The method comprises the following steps: firstly, collecting multi-source data such as historical load, renewable energy output, equipment operation state and maintenance record of a power grid, constructing a time sequence database through cleaning, time alignment and feature extraction, and establishing a load and renewable energy probability model; constructing a maintenance plan model containing a state variable, a constraint condition and a peak clipping weight mechanism, and establishing a continuous time Markov chain state model for the key equipment to generate an availability sequence; generating a large-scale random operation scene set through a Monte Carlo method based on multiple models, and carrying out supply-demand balance and power flow analysis on each scene; multi-dimensional indexes of reliability, economy and safety are calculated and subjected to weighted fusion, a comprehensive post-evaluation report is generated after results are counted, and finally the maintenance plan is optimized according to the report. According to the method, the uncertainty of the power system can be comprehensively considered, multi-dimensional quantitative evaluation and closed-loop optimization of the maintenance plan are realized, intelligent support is provided for power grid maintenance decision making, and the method is suitable for a power transmission network, a power distribution network and a micro-grid.

This invention provides a method for rapid assembly of a modular tubular beam car body. The car body is divided into several independent modules, which are manufactured in parallel at independent workstations or assembly lines. Corresponding connections between different independent modules are fitted with interlocking mortise and tenon structures for coarse positioning and fastening. The manufactured independent modules are then automatically transferred to an assembly station, where a robot grasps and rapidly assembles them. After assembly, bolts and / or structural adhesive are used to fasten the mortise and tenon joints. Compared to traditional assembly line production processes, this method significantly improves efficiency and economy, effectively reduces floor space, and overcomes the limitation of existing "unpacking processes" that are only suitable for single-model production, thus facilitating flexible manufacturing and rapid iteration across multiple models.

The invention discloses an optimal scheduling method for large model reasoning service and a related device. A plurality of reasoning service instances and a plurality of nodes used for running the reasoning service instances are deployed in the reasoning service of the target large model, firstly, hardware utilization indexes and service utilization indexes of all the reasoning service instances are collected, and collection time and instance IDs are recorded. Then, the indexes are preprocessed, and hardware preprocessing indexes and service preprocessing indexes are obtained; thirdly, multiple model parameters of the target large model and hardware computing power parameters of a target computer are obtained, and the ID, the hardware preprocessing index and the service preprocessing index of each reasoning instance are associated and integrated with the model and the computing power parameters to form an associated data set; and based on the data set, respectively calculating an HRUI, an SPHM and an HRVE of each reasoning service instance, so as to generate a capacity adjustment strategy and a request routing strategy of the target large model based on the HRUI, the SPHM and the HRVE.

The invention discloses a distribution box room environment parameter integrated measurement method, and particularly relates to the field of multi-parameter measurement, and the method comprises the steps: firstly constructing a machine room three-dimensional model and a sensor correlation degree matrix, then building a heat balance, humidity diffusion and airflow motion model, and calibrating parameters; deploying a measurement system and calibrating through dual synchronization and cross check; establishing a sensor confidence evaluation system and an adaptive threshold based on 72-hour reference data; fusing calibration data by adopting a three-level correlation calibration mechanism; four types of measurement modes and conversion logics are designed, and measurement resources are dynamically allocated through environmental risk assessment; multi-dimensional state parameters are extracted, a comprehensive evaluation value is calculated through normalization, dynamic weighting and combinatorial algorithms, and five-level early warning response is achieved; according to the method, multiple models and an intelligent algorithm are integrated, the parameter measurement precision and the environment risk identification efficiency are improved, the response delay is reduced, the fault diagnosis accuracy is improved, and reliable technical support is provided for safe operation and maintenance of the distribution box room.

The invention discloses a self-adaptive sensing anti-shake regulation and control method and system for a spinal surgery auxiliary mechanical arm, and relates to the technical field of spinal surgery, and the method comprises the steps: collecting multi-dimensional data of a mechanical arm tail end actuator through a multi-dimensional sensor group, and transmitting the multi-dimensional data to an AeroStead space-level stability calculation platform; the platform carries out filtering processing on the data, and vibration of the mechanical arm and disturbance components caused by external interference are separated; constructing a spine mechanical arm rigid-flexible coupling dynamic disturbance observation and compensation model by using the disturbance component to observe disturbance and generate a compensation amount; calling a spinal operation area three-dimensional reconstruction and dynamic obstacle avoidance mixed density network model, and fusing different images to generate an operation area three-dimensional dynamic model and an obstacle avoidance adjustment signal; a mechanical arm motion control instruction is generated through a force sense position cooperation self-adaptive control algorithm; and the mechanical arm is driven to move for closed-loop regulation and control. According to the method, through fusion of multiple models and algorithms, the disturbance compensation precision, the operation area modeling accuracy and the control collaboration are improved, and the operation safety is guaranteed.

The invention discloses a total primary productivity estimation method and system based on multi-model coupling deep learning, and the method comprises the steps: obtaining the multi-source data of meteorological data, remote sensing images, latent heat flux, sensible heat flux and solar radiation, carrying out the quality control, missing value processing and nearest neighbor interpolation of different data sources, and carrying out the prediction of the total primary productivity. Unifying to a target spatial resolution and a time resolution; in a light energy utilization rate (LUE) model family, a solar radiation phase factor is introduced into a photosynthetically active radiation absorption ratio (FPAR) to obtain a phase modulation type FPAR, drought duration is introduced into a water stress function f (W) to obtain an exponential decay type f (W), and a GPP time sequence of a plurality of improved mechanism models is calculated according to the exponential decay type f (W); extracting spatial texture features from a remote sensing image stack by using a convolutional neural network (CNN), and performing cross-modal fusion on the spatial features and the GPP estimated by the plurality of improved mechanism models in a gating mode to form fusion representation; carrying out learning and collaborative optimization on the fusion representation of the GPP and CNN spatial features estimated by the plurality of improved mechanism models by adopting a gradient lifting tree model; and high-precision estimation of the GPP is realized through multi-model collaborative optimization. The method aims at solving the problem that a traditional light energy utilization rate model is insufficient in response under the extreme environment conditions of drought and intense radiation, the adaptability limitation of a traditional single model under the complex environment is broken through, and therefore high-precision GPP estimation under the complex environment is achieved.

The invention discloses a GNN-PINN hybrid architecture-based propeller wake flow pressure field reconstruction device and method and a storage medium, and belongs to the field of flow field measurement. The method comprises the following steps: firstly, obtaining CFD simulation data of multi-model propellers under different working conditions and constructing a data set; adding noise to the velocity field data to simulate experimental noise; a hybrid architecture pressure reconstruction model based on deep fusion of a graph neural network (GNN) and a physical information neural network (PINN) is constructed, and the physical rationality of output is ensured through a physical constraint mechanism; after the network training is finished, PIV experimental data is used as a test set for verification, and the result shows that the GNN-PINN hybrid model can realize rapid and high-precision reconstruction of the wake flow pressure field. The method overcomes the defects of interference of traditional pressure measurement on a flow field and long consumed time of a CFD method, is suitable for wake flow pressure field reconstruction of propellers of various models and under various working conditions, and has high robustness and engineering application value.

The invention discloses an intelligent cue word generation evaluation method and system based on multi-component collaboration, and relates to the technical field of artificial intelligence and natural language processing. The method comprises the following steps: acquiring a multi-format text, converting the multi-format text into an input-output pair, and generating a plurality of candidate cue words; after determining a communication mode according to the text form, standardizing cue words according to the structured template of each large language model; calling each platform to generate answer analysis and unify formats, and extracting target contents by utilizing regularization; and carrying out multi-mode evaluation on the candidate cue words, selecting an optimal cue word, and outputting a structured evaluation result and an optimization suggestion based on multiple indexes. According to the method, full-process automatic processing of prompt word generation, optimization and evaluation is achieved, various model platforms, data sets, algorithms and evaluation indexes are supported, and good universality and expansibility are achieved.

The invention discloses a drainage pipe network data cleaning and intelligent repairing method fusing multiple models. The method comprises the following steps that S1, original monitoring data flow is collected; s2, carrying out the preliminary detection and elimination of the abnormity based on IQR; s3, performing upstream and downstream multi-source feature extraction and XGBoost weight analysis; step S4, multivariable depth autoregression prediction based on PatchTST is carried out; s5, generating and publishing continuous high-quality data; according to the method, the key technical problems of various abnormal types, incomplete abnormal detection, low cleaning and repairing precision, poor data continuity and trend consistency, high system integration and real-time processing difficulty and the like commonly existing in the actual collection and transmission process of the online monitoring data of the drainage pipe network are solved.

The invention discloses an earth surface downlink short-wave radiation space downscaling method based on a super-resolution reconstruction technology, belongs to the technical field of meteorological data analysis, and solves the problems of insufficient solar radiation data space resolution and weak earth surface heterogeneity characterization capability in the prior art. Carrying out super-resolution reconstruction on the radiation data set by adopting a super-resolution combination model, and carrying out weighted fusion on a reconstructed data set output by the super-resolution combination model; according to the method, the output results of the multiple models are subjected to weighted fusion, weight optimization of the super-resolution sub-models is performed in combination with the high-resolution terrain factors, spatial downscaling from 5km to 1km resolution is finally realized, the spatial precision of the solar radiation data is remarkably improved on the basis of keeping the hour-level time resolution of the original data, and the accuracy of the solar radiation data is improved. And more earth surface heterogeneity information can be captured, and meanwhile, the high-time-resolution application requirement is met.

The invention belongs to the technical field of electric power information physical system security, and discloses an urban power grid information physical system security situation early warning method based on cross-space risk propagation. A power grid physical layer and information layer coupling model and a cellular space false data injection attack model are constructed, and a fault cross-space propagation mechanism is simulated through an event-driven cellular automaton theory; an integrated kernel extreme learning machine model is adopted to predict the operation state of the power grid, multi-dimensional data fusion is realized through a radial basis kernel function, and an integrated structure is adopted to fuse a plurality of model prediction results so as to improve the model prediction precision; and establishing an early warning system including voltage out-of-limit, line overload and load loss indexes, and dynamically distributing weights and dividing early warning grades in combination with an entropy weight method. According to the method, the problem of low precision of urban power grid security situation early warning under multivariate disturbance is solved, and the accuracy and robustness of security situation early warning can be effectively improved.

A model-as-a-service (MaaS) platform performs cross-model resources allocation from a shared pool of GPU resources based on model-agnostic metrics generated by a metric standardizer. The metric standardizer receives, from model providers, model-specific benchmark metrics that define relationships between resource utilization and token processing according to the different model-specific tokenization schemes; receives, from one or more MaaS components, token-based job metrics pertaining to LLM processing tasks; and determines, based on the model-specific benchmark metrics and token-based job metrics, the model-agnostic metrics for multiple model pools executing instances of different large language models (LLMs) that generate and process text according to different model-specific tokenization schemes. The MaaS platform further includes one or more resource allocation components that dynamically reallocates resources of the shared pool based on the model-agnostic metric.

The application discloses a multi-model adaptive low-temperature valve sealing performance testing device and method, and belongs to the technical field of valve detection. The device comprises: a first upper end cover of a low-temperature thermostat, wherein a to-be-tested low-temperature valve is installed on the first upper end cover through a valve support device; the valve support device is in flexible sealing cooperation with a bellows through an adjusting screw, so that the installation height and centering of the to-be-tested low-temperature valve are controlled, and low-temperature valves of different specifications are adapted; the low-temperature thermostat surrounds a first cold screen and a second cold screen; the bottom of the to-be-tested low-temperature valve is connected with one end of a cold-lead bridge through a pipe clamp and is surrounded by the first cold screen; a GM refrigerating machine comprises two-stage cold heads arranged in the low-temperature thermostat; the second cold screen surrounds the two-stage cold heads; the other end of the cold-lead bridge is connected with the two-stage cold heads; three helium supply devices are connected with the inlet and outlet of the to-be-tested low-temperature valve and the low-temperature thermostat respectively; and a vacuum pump unit is connected with the low-temperature thermostat. The application realizes rapid adaptation and accurate testing of low-temperature valves of multiple models.

A data storage device can generate a recommended read threshold value using a model that was trained under a plurality of conditions. However, such a model may result in an undesirable bit error rate or programming latency. If that should occur, the data storage device can use a different model trained under a condition similar to a current condition of the data storage device. In addition to avoiding an undesirable bit error rate or programming latency, this can result in improved throughput, improved quality of service, and reduced power consumption.

The invention discloses a transformer temperature rise evaluation method and system based on a Stacking integrated learning framework, and the method comprises the steps: constructing a sample data set based on input variables including the size of an oil baffle plate, the width of an oil duct, the number of oil baffle plates and boundary temperature and response variables including the temperature rise of winding wire oil and the temperature rise of a hot spot through employing a Latin hypercube sampling principle and CFD simulation; constructing a Stacking integrated temperature rise model fusing the multi-model information in a layered manner, independently training a first layer by using the sample data set to generate a plurality of prediction results, and training a second layer after splicing the prediction results to obtain the Stacking integrated temperature rise model fusing the multi-model information; and after hyper-parameter joint optimization is carried out to determine hyper-parameters, a machine learning library is called to automatically complete training of the Stacking integrated temperature rise model fused with multi-model information, a trained temperature rise prediction model is obtained, temperature rise prediction is carried out on a to-be-predicted transformer configured by a new oil baffle plate and an oil duct structure, and predicted values of output line oil temperature rise and hot spot temperature rise values are obtained.

The invention provides a clearance measuring device for a small bearing. The clearance measuring device comprises a first accessory group and a second accessory group, the second accessory group comprises a plurality of accessory units corresponding to the model of the bearing to be detected; the first accessory group comprises a main gear, a fan-shaped rack meshed with the main gear, a dial indicator and a positioning assembly; by adopting the combination of the first accessory group and the second accessory group, the corresponding second accessory group can be selected and conveniently replaced according to the model of the bearing to be measured, so that one device can quickly adapt to different standard measuring forces required by various models of bearings; through meshing of the main gear and the fan-shaped rack, it is guaranteed that the pressure applied to the bearing is constant and can be repeated; the pressure value can be accurately set by replacing a force application spring and limiting of a corresponding limiting bolt and a stop pin, the standard measuring force requirements required by different bearing models are strictly met, and the accuracy and the reliability of a measuring result are fundamentally ensured.

The invention discloses an intelligent character recognition and structured processing method based on multi-modal deep learning, and belongs to the technical field of character recognition, and the method comprises the following steps: 1, receiving an image uploaded by a user, carrying out the detection of a character main body region, and extracting a main body region image containing characters; step 2, preprocessing the image; 3, performing character target detection on the preprocessed image by adopting a YOLOv8 model, and positioning a text region; step 4, identifying the detected character area by using a CRNN model, and outputting text content; 5, performing layout analysis on the recognition result, and determining a reading sequence and a logic structure of the text; and step 6, using a natural language model, combining the text content and the coordinate information, and carrying out structured output on an identification result. Multiple tasks such as business licenses and licenses are integrated into one model, the problem of multiple models in the prior art is solved, the resource utilization rate is high, and the recognition precision is high.

The invention provides a streaming output method of a model and electronic equipment, and is applied to the technical field of artificial intelligence, the method comprises the following steps: in response to a plurality of models, receiving a target problem sent by a user side, and respectively creating a corresponding event loop for each model in an asynchronous thread; executing an asynchronous streaming output task of the corresponding model based on an event loop in the asynchronous thread, and generating streaming data for the target problem; writing the streaming data output by each asynchronous thread into a message queue, wherein each streaming data has different first priorities for executing the asynchronous streaming output task based on the corresponding model; in response to the situation that the streaming data in the message queue reaches a preset capacity threshold value, determining target streaming data of which the first priority is lower than a preset priority threshold value from the message queue, and merging at least two lexical elements in the target streaming data; and obtaining lexical elements from the merged message queue, and sequentially returning the lexical elements to the user side in a synchronous stream output mode.

The invention relates to the technical field of artificial intelligence, can be applied to business scenes of financial science and technology, medical health and the like, and discloses a cross-model knowledge editing and updating method, device, equipment and medium, and the method comprises the steps: receiving a knowledge updating instruction, and generating a knowledge editing vector; writing the knowledge editing vector into an intermediate layer parameter of the target model set to obtain a parameter-updated target model set; generating an input sequence with a regulation mark when the input request and the knowledge update content meet semantic matching conditions, and inputting the input sequence into the target model set to obtain a model output set; performing output alignment and fusion through a consistency fusion module to generate a fusion output result; and when the output conflict is not eliminated, performing a re-reasoning operation based on the knowledge editing vector, and outputting a final result. According to the method, knowledge sharing and dynamic triggering among multiple models are realized through a cross-model mapping and consistency fusion mechanism, and the efficiency, timeliness and credibility of knowledge updating are improved.

An early screening system for chronic obstructive pulmonary disease based on blood routine test data. It belongs to the technical field of digital medical treatment, and specifically relates to the technical field of intelligent early screening of chronic obstructive pulmonary disease. It solves the technical problem that the generalization ability of a COPD prediction model is limited due to differences between different research data sets. The system comprises a data collection module that collects blood routine test data of a plurality of chronic obstructive pulmonary disease patients and blood routine test data of a plurality of normal personnel; a data preprocessing module that performs standardization processing on the blood routine test data to obtain a standardized data set; a screening model training module that trains N classification models using the standardized data set and calculates the prediction accuracy of each classification model; and a screening model application module that obtains the fusion weight of each model by using a normalization processing method based on the prediction accuracy of each classification model, fuses the risk scores predicted by the multiple models, and obtains a final screening result.

The invention discloses a steel casting grinding method and device based on 3D virtual-real combination man-machine cooperation. Firstly, casting three-dimensional point cloud is obtained, a polishing path is generated based on the casting three-dimensional point cloud, and then simulated polishing is conducted based on the polishing path. And the worker judges whether the polishing path is reasonable or not based on the casting three-dimensional point cloud and the target three-dimensional model, if not, the polishing path is corrected, and simulated polishing is repeated till the polishing path is reasonable. According to the method, the grinding path can be generated according to the three-dimensional point cloud of the steel casting, the grinding path can better fit the situation of the steel casting, and the machining requirements of castings of various models and sizes and different machining precision requirements can be met. And meanwhile, polishing simulation is further carried out, the reasonability of the polishing path is judged according to the effect of the target three-dimensional model, and the situation that the polishing robot carries out polishing based on the unreasonable polishing path is avoided. And compared with a fixed path grinding mode, the precision and the generation efficiency of steel casting grinding can be greatly improved.

The invention provides a version comparison method and device of a service model and a storage medium. The method comprises the following steps: receiving a comparison request sent by a client to the service model; querying model object instances of the business model under multiple versions according to the comparison request; annotation information is added to metadata in the model object instance; comparing metadata in the plurality of model object instances according to the annotation information to obtain version difference information; and sending the version difference information to the client for display. According to the embodiment, based on a differentiation comparison strategy of annotation information, the object hierarchy and the set element structure with the complex nested relationship in the business model can be deeply analyzed, the change of the attribute value can be detected, deep structure differences such as the change of the object reference relationship and the logic increase and decrease of the set element can be more accurately identified, and the service performance of the business model is improved. And comprehensive and accurate model change information is provided for the user.

The invention provides a database abnormity intelligent detection method and system based on multi-source heterogeneous data fusion, and relates to the technical field of database management, and the method comprises the following steps: obtaining key data according to original data of a database; extracting an association relationship between entities from the key data, and adding corresponding timestamps to the entities and the association relationship; arranging the entities according to the time stamp sequence and associating the entities with the association relationship to obtain a multi-dimensional time knowledge graph; decomposing the multi-dimensional time knowledge graph into a plurality of static knowledge graphs with a time sequence relationship, and obtaining a time sequence according to the association relationship between entities in the plurality of static knowledge graphs; adopting a feature fusion model to obtain a feature vector corresponding to each entity based on the time sequence; and adopting a time sequence detection model to identify abnormal behaviors based on the feature vector corresponding to each entity. According to the method, accurate detection of the abnormal behavior of the database is realized by combining the multi-source data with multiple models.

Systems, methods, and devices that relate to generation of responses to query sets are disclosed. In one example aspect, the system uses multiple models to retrieve data relevant to queries and appropriate for the requesting user and to output the data in a certain style. In particular, the system uses a first model to determine a type of user submitting a query, a second model to retrieve data relevant to the query, subject to constraints based on the type of user, and a third model to formulate a response to the query, based on the retrieved data, using a consistent style. In some implementations, another model enables certain users to validate the outputs from one or more of the first, second, and third models. The system can output a compilation of the responses to the queries in a particular manner, style, or presentation that is consistent across all outputs.

The invention discloses a ChatBI plug-in construction method and system based on multiple models. The method comprises the steps that a data source is established, and configuration connection is carried out; selecting successfully connected data sources and databases and tables under the data sources to form a database mode; selecting a pre-integrated natural language processing model type and inputting an API-key to complete information registration of the natural language processing model; the pre-integrated natural language processing model comprises a plurality of models; aiming at different SQL dialogue scenes, configuring corresponding prompt word templates; aiming at different SQL dialogue scenes, configuring corresponding dialogue interaction strategies; configuring application basic information, and selecting a corresponding database mode, a model, a cue word template and a dialogue interaction strategy according to the application basic information to form a query scene application; and generating a ChatBI plug-in package according to the query scene application. The problems that a traditional BI system lacks intelligence and is inconvenient to integrate with an original system are solved.

A model-as-a-service (MaaS) platform performs cross-model resources allocation from a shared pool of GPU resources based on model-agnostic metrics generated by a metric standardizer. The metric standardizer receives, from model providers, model-specific benchmark metrics that define relationships between resource utilization and token processing according to the different model-specific tokenization schemes; receives, from one or more MaaS components, token-based job metrics pertaining to LLM processing tasks; and determines, based on the model-specific benchmark metrics and token-based job metrics, the model-agnostic metrics for multiple model pools executing instances of different large language models (LLMs) that generate and process text according to different model-specific tokenization schemes. The MaaS platform further includes one or more resource allocation components that dynamically reallocates resources of the shared pool based on the model-agnostic metric.