A large model intelligent migration system, method, device and medium

Abstract

The application discloses a large model intelligent migration system, method, device and medium, comprising: searching and judging whether there is a completely matched target historical successful execution experience in a script experience knowledge base, in the case of hitting the target historical successful execution experience, generating a target migration parameter configuration scheme for the current model migration task according to the target historical successful execution experience, in the case of not hitting the target historical successful execution experience, fusing the retrieved candidate migration script, historical failed migration experience associated with the current model migration task, architecture features of the model to be migrated and computing power conditions of the target migration hardware to generate the target migration parameter configuration scheme, collecting execution state and log information of the current model migration task, and generating optimization suggestions according to the execution state, log information and execution effect after completion, and storing the target migration parameter configuration scheme and optimization suggestions into the script experience knowledge base to generate new historical execution experience.

Description

Technical Field

[0001] This application belongs to the field of model transfer technology, specifically relating to a large model intelligent transfer system, method, device and medium. Background Technology

[0002] In recent years, with the rapid development of large-scale language model technology, large models based on Transformer have achieved remarkable results in natural language understanding, content generation, and intelligent dialogue. Dedicated AI computing chips have gradually become a key infrastructure supporting the efficient operation of large models. Among them, the NPU (neural processing unit), with its self-developed Da Vinci architecture and comprehensive tool ecosystem, occupies an important position in AI computing platforms and is a crucial hardware carrier for the efficient inference and deployment of large language models. Currently, migrating and adapting large language models from traditional GPU (graphics processing unit) platforms to NPU platforms is of significant practical importance for reducing deployment costs and promoting the development of the computing ecosystem.

[0003] The overall process of this model migration can be roughly divided into three core stages: weight transformation and adaptation, accuracy debugging, and pre-training iterative verification, involving a series of tasks such as model parsing and inference verification. Different large language models differ in structure, number of parameters, and operator types, resulting in highly diverse and scenario-dependent parameter configurations and script calls at each migration stage. Currently, migration work mainly relies on manual debugging and fixed interface development. However, this method is not only cumbersome and inefficient, but also leads to poor system reusability and difficulty in expansion due to the need to develop separate adaptation interfaces for different models. Furthermore, migration operations require a high level of expertise from users, making it difficult for ordinary users to complete migration tasks quickly and stably.

[0004] Against this backdrop, the cross-hardware migration process of large language models faces challenges such as complex procedures, numerous parameters, difficulty in reusing experience, and high user barriers. Summary of the Invention

[0005] The purpose of this application is to provide a large-scale intelligent migration system, method, device, and medium that can solve the above-mentioned problems.

[0006] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a large-model intelligent transfer system, which includes at least: A script experience knowledge base is used to store the historical execution experience and migration scripts of multiple historical model migration tasks; the historical execution experience includes historical successful execution experience and historical failed execution experience. The intelligent decision-making layer, when a current model migration task is initiated, uses script tools to search and determine whether there is any historical successful execution experience that perfectly matches the current model migration task in the script experience knowledge base, and obtains the search results: The intelligent decision-making layer is used to generate a target migration parameter configuration scheme for the current model migration task according to the target's historical successful execution experience when the search results indicate that there is such experience. The intelligent decision layer is used to, when the search results indicate that there is no successful historical execution experience of the target, integrate the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware, to generate a target migration parameter configuration scheme. The intelligent decision-making layer is used to collect the execution status, log information, and execution effect of the current model migration task using monitoring tools, and after migrating the model to be migrated to the target migration hardware according to the target migration parameter configuration scheme, it generates optimization suggestions based on the execution status, log information, and execution effect of the current model migration task. The intelligent decision-making layer is used to utilize the monitoring tool to generate new historical execution experience from the target migration parameter configuration scheme and the optimization suggestions and store it in the script experience knowledge base.

[0007] Optionally, the step of fusing the candidate migration scripts retrieved by the script tool from the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool from the script experience knowledge base, with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme includes: The intelligent decision layer is used to receive the script tool generating search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated, and to determine the candidate migration scripts from the script experience knowledge base based on the search keywords. The intelligent decision-making layer is used to receive historical failure migration experiences determined by the memory tool from the script experience knowledge base based on the search keywords; The intelligent decision-making layer is used to construct a basic parameter set based on the parameter names and default values ​​extracted from the candidate migration script; The intelligent decision-making layer is used to generate a target migration parameter configuration scheme based on the historical failed migration experience, the basic parameter set, the computing power information of the target migration hardware, and the architectural characteristics of the model to be migrated.

[0008] Optionally, the script tool generates search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model's unique identifier, and, based on the search keywords, determines candidate migration scripts from the script experience knowledge base, including: A script tool is used to obtain the configuration file of the model to be migrated and parse it to obtain the architectural features of the model to be migrated; A scripting tool is used to generate search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model's unique identifier. Based on these search keywords, it then initially filters out multiple candidate migration scripts that satisfy the search keywords from the script experience knowledge base. A script tool is used to determine the first model series information and the first model size information of the model to be migrated based on the configuration file of the model to be migrated, and to determine the second model series information and the second model size information of the model to be migrated based on the unique identifier of the model to be migrated. A script tool is used to compare the first model series information and the first model size information with the second model series information and the second model size information to obtain a script filtering range, and to filter the multiple initial candidate migration scripts based on the script filtering range to obtain multiple intermediate candidate migration scripts. A scripting tool is used to determine candidate migration scripts from a plurality of intermediate candidate migration scripts by utilizing the first major language model, based on the computing power information of the target migration hardware, the functional type of the model to be migrated, and the configuration file of the model to be migrated.

[0009] Optionally, the historical failed migration experiences determined from the script experience knowledge base based on the search keywords include: An original query is generated based on the search keywords, the migration steps, and the configuration information. The original query is then semantically expanded using the second major language model to obtain an expanded query. Based on the original query and the extended query, multiple initial candidate failed migration experiences are determined from the script experience knowledge base; By using a weighted fusion ranking method, based on the vector semantic similarity and time decay factor between the current model transfer task and the initial candidate failed transfer experiences, the top K initial candidate failed transfer experiences are determined as historical failed transfer experiences.

[0010] Optionally, the current model transfer task includes: weight transformation, model training, accuracy acquisition, and performance acquisition. The system also includes: The interactive access layer is used to receive the unique identifier of the model to be migrated from user input. A service collaboration layer is used to download the model to be migrated based on the model's unique identifier. The intelligent decision-making layer is used to obtain the architectural features of the model to be migrated and the migration steps of the model to be migrated based on the model to be migrated downloaded by the service collaboration layer. The interactive access layer is used to display the target migration parameter configuration scheme generated by the intelligent decision layer to the user, and to receive the user's confirmation information or parameter configuration adjustment information on the target migration parameter configuration scheme in the interactive access layer. The intelligent decision-making layer, upon receiving the parameter configuration adjustment information, combines the parameter configuration adjustment information, the basic parameter set, the historical failure execution experience, the computing power information of the target migration hardware, the architectural features of the model to be migrated, and preset decision guidance prompts. It then optimizes the basic parameter set through a third language model to obtain an updated target migration parameter configuration scheme, which is displayed to the user through the interactive access layer until the user confirms the updated target migration parameter configuration scheme through the interactive access layer. The computation execution layer is used to execute the current model migration task according to the target migration parameter configuration scheme or the updated target migration parameter configuration scheme after receiving the confirmation information; The migration script in the target migration parameter configuration scheme or the migration script in the updated target migration parameter configuration scheme includes the model format conversion script required for weight conversion, the forward verification script for the target migration hardware, and the model training script required for model training.

[0011] Optionally, the computation execution layer includes: a weight transformation module, a model training module, an accuracy acquisition module, and a performance acquisition module; executing the current model transfer task according to the target transfer parameter configuration scheme or the updated target transfer parameter configuration scheme includes: The weight conversion module is used to sequentially perform format conversion, weight consistency verification, original migration hardware forward verification, target migration hardware forward verification, and forward result comparison on the model to be migrated according to the model format conversion script and the forward verification script of the target migration hardware. The model training module is used to perform data preprocessing and pre-training iterations on the model to be transferred, according to the model training script. The accuracy acquisition module is used to provide the user with training process accuracy data when the user enables the training data acquisition function before the model to be transferred is pre-trained. The performance acquisition module is used to provide users with training process performance data when the user enables the training data acquisition function before the model to be transferred is pre-trained.

[0012] Optionally, the interactive access layer is also used to provide a custom script configuration entry point for models that have not yet been adapted; The custom script configuration entry is used to allow users to write and submit migration scripts for models that have not yet been adapted, and to store the migration scripts for models that have not yet been adapted in the script experience knowledge base for retrieval and reuse by subsequent model migration tasks of the same type.

[0013] Secondly, embodiments of this application provide a large-model intelligent transfer method, the method comprising: The historical execution experience and migration scripts of multiple historical model migration tasks are stored in the script experience knowledge base; the historical execution experience includes historical successful execution experience and historical failed execution experience. When the current model migration task is initiated, the script tool is used to search and determine whether there is any historical successful execution experience that completely matches the current model migration task in the script experience knowledge base, and the search results are obtained: If the search results indicate the existence of successful historical execution experience for the target, a target migration parameter configuration scheme for the current model migration task is generated according to the successful historical execution experience for the target. If the search results indicate that there is no successful historical execution experience of the target, the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, are fused with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme. The execution status, log information, and execution effect of the current model migration task are collected using monitoring tools. After the model to be migrated is migrated to the target migration hardware according to the target migration parameter configuration scheme, optimization suggestions are generated based on the execution status, log information, and execution effect of the current model migration task. Using the monitoring tool, the target migration parameter configuration scheme and the optimization suggestions are used to generate new historical execution experience and stored in the script experience knowledge base.

[0014] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps in the large model intelligent migration method as described in the second aspect.

[0015] Fourthly, embodiments of this application provide a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps in the large model intelligent migration method as described in the second aspect.

[0016] This application provides a large-scale intelligent migration system. The system includes at least: a script experience knowledge base for storing historical execution experiences and migration scripts for multiple historical model migration tasks; the historical execution experiences include historical successful execution experiences and historical failed execution experiences; and an intelligent decision layer for, when a current model migration task is initiated, using script tools to search and determine whether a target historical successful execution experience that perfectly matches the current model migration task exists in the script experience knowledge base, obtaining a search result: if the search result indicates the existence of a target historical successful execution experience, a target migration parameter configuration scheme for the current model migration task is generated according to the target historical successful execution experience; if the search result indicates the absence of a target historical successful execution experience, the script... The tool retrieves candidate migration scripts from the script experience knowledge base, and the memory tool retrieves historical failed migration experiences associated with the current model migration task from the script experience knowledge base. These are then fused with the architectural characteristics of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme. A monitoring tool collects the execution status, log information, and execution results of the current model migration task. After migrating the model to be migrated to the target migration hardware according to the target migration parameter configuration scheme, optimization suggestions are generated based on the execution status, log information, and execution results of the current model migration task. Finally, the monitoring tool uses the target migration parameter configuration scheme and the optimization suggestions to generate new historical execution experiences and stores them in the script experience knowledge base.

[0017] The technical solution of this application constructs a large-scale intelligent migration system. It addresses the problems of script selection relying on manual intervention, insufficient flexibility in parameter adaptation, and difficulty in reusing historical migration experience in traditional migration processes. The system achieves the matching of historical migration experience through a hierarchical collaborative architecture, and a closed-loop optimization mechanism that generates target migration parameter configuration schemes, task execution, and historical execution experience records by multi-source information fusion. This reduces the reliance on manual intervention in heterogeneous migration scenarios and improves the efficiency and stability of cross-hardware migration of large language models. Attached Figure Description

[0018] Figure 1 This is a comparison diagram of the traditional migration process provided in one embodiment of this application and the large-scale intelligent migration system of this application; Figure 2 This is an overall framework diagram of a large-scale intelligent transfer system provided in one embodiment of this application; Figure 3 This is an overall framework diagram of a multi-source information hybrid retrieval process in a large-scale intelligent migration system provided in an embodiment of this application; Figure 4 This is a flowchart illustrating a method for a large-scale intelligent transfer system provided in an embodiment of this application; Figure 5 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0021] Figure 1This diagram compares the traditional migration process provided in one embodiment of this application with the intelligent migration system for large models presented in this application. Although current agent technology has made significant progress in areas such as general tasks, code generation, and conventional algorithm optimization, there is still a lack of targeted agent design and decision-making mechanism research for cross-hardware migration of large language models, such as migration from GPU to NPU—a highly specialized scenario with strong hardware constraints and high parameter dependencies. Existing migration processes rely entirely on manually written scripts and manual task execution, lacking intelligent decision-making, experience reuse, and dynamic scheduling capabilities, resulting in high deployment barriers and low efficiency. Therefore, it is urgent to design an agent architecture and parameter recommendation mechanism specifically tailored to the unique characteristics of this scenario to address pain points such as cumbersome configuration and difficulty in reusing experience, thereby promoting the intelligent and large-scale deployment of large models on domestic computing platforms.

[0022] The following description, in conjunction with the accompanying drawings, details a large-model intelligent migration system, method, device, and medium provided in this application through specific embodiments and application scenarios.

[0023] The cross-hardware migration of large language models from GPUs to NPUs is a core step in deploying on heterogeneous computing platforms. It involves several key aspects, including environment adaptation, model weight format conversion, and parameter configuration, demanding a high level of expertise and experience from users. While the conventional manual migration method, based on official technical documentation, is technically standardized and feasible, it requires users to fully understand and independently complete a series of operations, including environment deployment, script selection, parameter modification, and command execution. This process is cumbersome and has a high learning curve, making it difficult to meet the practical needs for efficient migration of large language models.

[0024] Figure 2 This application provides a system architecture diagram of a large-scale intelligent migration system according to one embodiment; Figure 3 This is an overall framework diagram of a multi-source information hybrid retrieval process in a large-scale intelligent migration system provided in an embodiment of this application.

[0025] Based on this, one embodiment of this application provides a large-model intelligent transfer system, the overall architecture of which is as follows: Figure 2 As shown, a four-layer progressive design is adopted, consisting of an interaction access layer, a service collaboration layer, an intelligent decision-making layer, and a computing execution layer. Each layer distributes tasks and exchanges data via RESTful interfaces, and uses Server-Sent Events (SSE) long connections to achieve real-time push of runtime logs and task progress, forming an end-to-end migration loop from user operation to Ascend NPU execution. This architecture physically isolates the development environment and execution environment for cross-hardware migration, effectively avoiding dependency conflicts between different frameworks, while supporting elastic scaling and flexible scheduling of computing nodes in the computing cluster of the computing execution layer.

[0026] In one embodiment, the system includes at least: A script experience knowledge base is used to store the historical execution experience and migration scripts of multiple historical model migration tasks; the historical execution experience includes historical successful execution experience and historical failed execution experience. The intelligent decision-making layer, when a current model migration task is initiated, uses script tools to search and determine whether there is any historical successful execution experience that perfectly matches the current model migration task in the script experience knowledge base, and obtains the search results: The intelligent decision-making layer is used to generate a target migration parameter configuration scheme for the current model migration task according to the target's historical successful execution experience when the search results indicate that there is such experience. The intelligent decision layer is used to, when the search results indicate that there is no successful historical execution experience of the target, integrate the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware, to generate a target migration parameter configuration scheme. The intelligent decision-making layer is used to collect the execution status, log information, and execution effect of the current model migration task using monitoring tools, and after migrating the model to be migrated to the target migration hardware according to the target migration parameter configuration scheme, it generates optimization suggestions based on the execution status, log information, and execution effect of the current model migration task. The intelligent decision-making layer is used to utilize the monitoring tool to generate new historical execution experience from the target migration parameter configuration scheme and the optimization suggestions and store it in the script experience knowledge base.

[0027] The large-scale intelligent migration system adopts a four-layer progressive design: the interaction access layer, the service collaboration layer, the intelligent decision-making layer, and the computation execution layer. Each layer distributes tasks and exchanges data via RESTful interfaces, and uses Server-Sent Events (SSE) long-lived connections to push runtime logs and task progress in real time, forming an end-to-end migration loop from user operation to NPU execution. This architecture physically isolates the GPU development environment from the NPU execution environment, effectively avoiding dependency conflicts between different frameworks, while supporting elastic scaling and flexible scheduling of computing nodes. The specific design of each layer will be described in detail below.

[0028] In this embodiment, the intelligent decision-making layer is centered on a parameter recommendation agent, whose decision-making and evolution rely on the coordinated efforts of scripting tools, memory tools, and monitoring tools. The scripting tools are responsible for constructing the basic parameter set, the memory tools are responsible for retrieving historical execution experience, and the monitoring tools collect execution results and generate optimization suggestions. Together, these three constitute a complete closed loop from parameter decision-making to experience evolution.

[0029] To support parameter recommendation in the transfer scripts of the parameter recommendation agent, the system has constructed a script experience knowledge base covering mainstream large language models. This base gathers transfer scripts and historical execution experience of various models in the weight transformation and pre-training transfer steps, providing an initial experience foundation for script tool retrieval and agent recommendation decisions. The statistical information is shown in Table 1.

[0030] Table 1: Parameter statistics of migration scripts in the script experience knowledge base

[0031] Specifically, the script experience knowledge base contains historical execution experience and transfer scripts for multiple historical model transfer tasks. Transfer scripts refer to the scripts used in transfer steps such as weight transformation and pre-training. Historical execution experience includes historical successful execution experience and historical failed execution experience. Historical successful execution experience refers to the historical execution experience of successfully executed historical model transfer tasks, and historical failed execution experience refers to the historical execution experience of failed historical model transfer tasks.

[0032] In the actual model migration process, the first step is to obtain the current model migration task issued by the user. The current model migration task refers to the cross-hardware migration task of the model to be migrated (large language model) from the original migration hardware to the target migration hardware, such as the cross-hardware migration from GPU to NPU. The script tool of the parameter recommendation agent of the intelligent decision layer, combined with the relevant information of the current model migration task (such as user identifier, model name, function type, parameter keywords, etc.), performs a preliminary search in the script experience knowledge base containing multiple historical migration experiences to obtain the search results. Then, it is determined whether there is a historical successful execution experience that completely matches the current model migration task among the multiple historical successful migration experiences contained in the search results. If so, it is determined as the target historical successful execution experience. The target historical successful migration experience is the optimal parameter configuration scheme adapted to the current model migration task, which is generated and verified by multi-source information fusion in the previous historical model migration task. Then, the optimal parameter configuration scheme in the target historical successful execution experience is reused to generate the target migration parameter configuration scheme for the current model migration task.

[0033] refer to Figure 3If there is no completely matching target historical successful execution experience, then the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, are fused with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme for the current model migration task.

[0034] During the execution of the current model migration task according to the target migration parameter configuration scheme, the monitoring tool collects the execution status and log information of the current model migration task. After the model to be migrated corresponding to the current model migration task is migrated from the original migration hardware to the target migration hardware according to the target migration parameter configuration scheme, optimization suggestions are generated based on the execution status, log information and execution effect of the current model migration task.

[0035] Finally, monitoring tools are used to store the target migration parameter configuration scheme and optimization suggestions into the new historical execution experience generated and stored in the script experience knowledge base. The target migration parameter configuration scheme includes the migration script for the current model migration task.

[0036] Through the above embodiments, the system constructs a script experience knowledge base covering mainstream large language models. During decision-making, it prioritizes retrieving and reusing matching historical successful execution experiences. If a matching historical execution experience is found, the corresponding parameter configuration scheme is directly reused to quickly adopt the validated optimal solution, generate the target migration parameter configuration scheme, and execute it. When a completely matching historical successful execution experience is lacking, the system uses candidate migration scripts retrieved by the script tool as initial parameters, historical failed migration experiences associated with the current model migration task, the architectural features of the model to be migrated, and the computing power conditions of the target migration hardware to adaptively generate the target migration parameter configuration scheme. After each model migration, the system automatically collects execution results and generates optimization suggestions. These suggestions, combined with the target migration parameter configuration scheme, generate new historical execution experiences that are persistently stored in memory. This persistent storage of historical execution experiences drives continuous iteration of the recommendation strategy, achieving adaptive evolution that becomes more accurate with each use.

[0037] Optionally, in one embodiment, the step of fusing the candidate migration scripts retrieved by the script tool from the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool from the script experience knowledge base, with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme includes: The intelligent decision layer is used to receive the script tool generating search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated, and to determine the candidate migration scripts from the script experience knowledge base based on the search keywords. The intelligent decision-making layer is used to receive historical failure migration experiences determined by the memory tool from the script experience knowledge base based on the search keywords; The intelligent decision-making layer is used to construct a basic parameter set based on the parameter names and default values ​​extracted from the candidate migration script; The intelligent decision-making layer is used to generate a target migration parameter configuration scheme based on the historical failed migration experience, the basic parameter set, the computing power information of the target migration hardware, and the architectural characteristics of the model to be migrated.

[0038] In this embodiment, the script tool generates search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated (the model unique identifier MID provided by the user, for example, from the ModelScope or Hugging Face platform). Based on these search keywords, it determines candidate migration scripts from the script experience knowledge base. At the same time, the memory tool determines historical failed migration experiences related to the current model migration task from the script experience knowledge base based on the search keywords generated by the script tool. Then, the parameter recommendation agent of the intelligent decision layer receives the historical failed migration experiences retrieved by the memory tool and the candidate migration scripts retrieved by the script tool, extracts parameter names and default values ​​from the candidate migration scripts, constructs a basic parameter set, and then generates a target migration parameter configuration scheme based on the historical failed migration experiences, the basic parameter set, the computing power information of the target migration hardware, and the architectural features of the model to be migrated.

[0039] Optionally, in one embodiment, the scripting tool generates search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated, and determines candidate migration scripts from the script experience knowledge base based on the search keywords, including: A script tool is used to obtain the configuration file of the model to be migrated and parse it to obtain the architectural features of the model to be migrated; A scripting tool is used to generate search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model's unique identifier. Based on these search keywords, it then initially filters out multiple candidate migration scripts that satisfy the search keywords from the script experience knowledge base. A script tool is used to determine the first model series information and the first model size information of the model to be migrated based on the configuration file of the model to be migrated, and to determine the second model series information and the second model size information of the model to be migrated based on the unique identifier of the model to be migrated. A script tool is used to compare the first model series information and the first model size information with the second model series information and the second model size information to obtain a script filtering range, and to filter the multiple initial candidate migration scripts based on the script filtering range to obtain multiple intermediate candidate migration scripts. A scripting tool is used to determine candidate migration scripts from a plurality of intermediate candidate migration scripts by utilizing the first major language model, based on the computing power information of the target migration hardware, the functional type of the model to be migrated, and the configuration file of the model to be migrated.

[0040] In this embodiment, for cases where the target's historical successful migration experience is not matched, the process proceeds to the migration script selection stage of the script tool. This provides an initial parameter basis for the subsequent multi-source information fusion to generate the target migration parameter configuration scheme. The migration script selection stage includes: initial script screening and intelligent selection of multiple scripts. 1) Initial Screening of Migration Scripts: The script tool obtains the configuration file of the model to be migrated in the ModelScope or Hugging Face platform. Then, combining this configuration information with the parsed architectural features, migration steps, and unique model identifier of the model, it generates multi-dimensional search keywords. Based on these keywords, it initially screens out a set of initial candidate scripts that meet the criteria from the script experience knowledge base; these are multiple initial candidate migration scripts. If there is only one initial candidate migration script, that initial candidate migration script is used as the candidate migration script.

[0041] 2) When the initial screening results contain multiple initial candidate migration scripts, the script tool will introduce the internal details in the configuration file of the model to be migrated for multi-dimensional in-depth comparison.

[0042] First, core parameters such as `model_type`, `hidden_size` (hidden layer dimension), and `num_layers` (number of hidden layers) are extracted from the configuration file of the model to be transferred, determining the first model family information and the first model size information of the model to be transferred. Simultaneously, the script tool extracts the second model family information and the second model size information from the model's unique identifier, and cross-validates and integrates them with the first model family information and the first model size information in the configuration file to obtain a more precise script selection range. Based on this script selection range, multiple initial candidate transfer scripts are filtered to obtain multiple intermediate candidate transfer scripts. If there is only one intermediate candidate transfer script, that intermediate candidate transfer script is selected as the final candidate transfer script.

[0043] When multiple intermediate candidate transfer scripts exist, the script tool will call the pre-trained first language model, and based on the computing power information of the target transfer hardware, the functional type of the model to be transferred, the configuration file of the model to be transferred, and the pre-designed prompt words, determine the intermediate candidate transfer script with the highest fit from the multiple intermediate candidate transfer scripts as the candidate transfer script.

[0044] Optionally, in one embodiment, the step of determining historical failed migration experiences from the script experience knowledge base based on the search keywords includes: An original query is generated based on the search keywords, the migration steps, and the configuration information. The original query is then semantically expanded using the second major language model to obtain an expanded query. Based on the original query and the extended query, multiple initial candidate failed migration experiences are determined from the script experience knowledge base; By using a weighted fusion ranking method, based on the vector semantic similarity and time decay factor between the current model transfer task and the initial candidate failed transfer experiences, the top K initial candidate failed transfer experiences are determined as historical failed transfer experiences.

[0045] In this embodiment, the historical execution experience retrieval stage is handled by a memory tool. The goal is to retrieve multiple historical failed execution experiences and their optimization suggestions that are highly correlated with the scenario of the current model transfer task from the vector database. This helps the parameter recommendation agent avoid known errors in historical failed execution experiences and provides decision-making references for subsequent multi-source information fusion.

[0046] The historical execution experience retrieval phase includes two core components: multi-query expansion and weighted hybrid sorting. 1) Multi-query expansion: The original query Q is constructed using the search keywords generated during the migration script screening stage, the name of the current migration step, and the model configuration information. The pre-trained second language model performs multi-faceted semantic expansion on the original query Q, generating several semantically equivalent but perspective-differentiated expanded queries. By integrating the original query and multiple expanded queries, a wider-coverage retrieval query set is constructed, improving the recall and matching completeness of relevant historical execution experience.

[0047] 2) Weighted Hybrid Ranking: After completing multi-query extended retrieval using the retrieval query set, multiple initial historical failure execution experiences are obtained. Then, the memory tool summarizes all matched initial historical failure execution experiences, introduces vector semantic similarity and time decay factor for multi-dimensional weighted fusion calculation, and performs global ranking of the initial historical failure execution experiences in the retrieval results. Priority is given to retaining initial historical failure execution experiences with high semantic matching degree and strong timeliness. Finally, the top-K highly correlated initial historical failure execution experiences are selected as historical failure migration experiences, providing reliable historical data support for subsequent multi-source information fusion.

[0048] Optionally, in one embodiment, the current model transfer task includes: weight transformation, model training, accuracy acquisition, and performance acquisition; the system further includes: The interactive access layer is used to receive the unique identifier of the model to be migrated from user input. A service collaboration layer is used to download the model to be migrated based on the model's unique identifier. The intelligent decision-making layer is used to obtain the architectural features of the model to be migrated and the migration steps of the model to be migrated based on the model to be migrated downloaded by the service collaboration layer. The interactive access layer is used to display the target migration parameter configuration scheme generated by the intelligent decision layer to the user, and to receive the user's confirmation information or parameter configuration adjustment information on the target migration parameter configuration scheme in the interactive access layer. The intelligent decision-making layer, upon receiving the parameter configuration adjustment information, combines the parameter configuration adjustment information, the basic parameter set, the historical failure execution experience, the computing power information of the target migration hardware, the architectural features of the model to be migrated, and preset decision guidance prompts. It then optimizes the basic parameter set through a third language model to obtain an updated target migration parameter configuration scheme, which is displayed to the user through the interactive access layer until the user confirms the updated target migration parameter configuration scheme through the interactive access layer. The computation execution layer is used to execute the current model migration task according to the target migration parameter configuration scheme or the updated target migration parameter configuration scheme after receiving the confirmation information; The migration script in the target migration parameter configuration scheme or the migration script in the updated target migration parameter configuration scheme includes the model format conversion script required for weight conversion, the forward verification script of the original migration hardware, and the model training script required for model training.

[0049] In this embodiment, the interaction access layer serves as the entry point for operation between the user and the system. It is a visual interactive interface built using the Vue3+TypeScript+Vite technology stack. This layer supports the unique model identifier corresponding to the user-input model to be migrated on the HuggingFace or ModelScope platform.

[0050] The service collaboration layer, built on Django and the Django REST Framework, is the main layer carrying the system's business logic. It is responsible for core services such as user authentication, permission management, model migration task distribution, and task execution status synchronization. This layer acts as a bridge between the interaction access layer and the computation execution layer, handling business requests from the interaction access layer. It only forwards computationally intensive tasks directly related to the target migration model to the computation execution layer via HTTP / RESTful interfaces; all other business logic is completed within the service collaboration layer.

[0051] The service collaboration layer receives a unique model identifier from the interaction access layer and downloads the model to be migrated based on this identifier (e.g., downloading the weights and architecture information of the model to be migrated), eliminating the need for manual model downloading in the traditional migration process. For self-developed models that are fine-tuned based on mainstream language models supported by the system, users only need to place the model files of such models to be migrated in the directory specified by the service collaboration layer and follow the unified naming convention, and the system will automatically recognize and integrate them without any additional interface adaptation work.

[0052] In terms of process guidance, the interactive access layer divides the current model migration task into clear step navigation, with each step corresponding to an independent parameter configuration page. Users follow the interface guidance to enter each step page step by step. The intelligent decision layer generates the corresponding target migration parameter configuration scheme based on the architectural characteristics of the model to be migrated, the migration steps (such as the name information of the migration steps), and the computing power information of the target hardware. The interactive access layer dynamically renders the adapted interactive configuration form based on the generated target migration parameter configuration scheme. After the user verifies the parameters in the interactive configuration form, they click the execute button to generate confirmation information, and the current model migration task can be initiated.

[0053] Meanwhile, the system allows users to manually adjust parameter configurations at the interactive access layer according to their actual needs for the target migration parameter configuration scheme, and generate parameter configuration adjustment information, including but not limited to parallel strategy adjustment, total number of iterations for model training adjustment, and train-iters adjustment.

[0054] The user-adjustable parameter configurations and the reasons for modification are as follows: `seq-length` and `max-position-embeddings` control the length of input text that the transferable model can handle. Increase these values ​​to accommodate the full context when processing long documents or multi-turn dialogues; decrease them to reduce memory overhead in short question-answering or classification tasks. They are typically kept consistent.

[0055] Tensor parallelism (TP) and pipeline parallelism (PP) are used to adapt to the hardware specifications and parallelism strategies of the computing cluster in the computation execution layer. When a single GPU memory overflows, increasing TP distributes the model parameters in a single computation across more GPUs to reduce the memory usage of a single GPU. When cross-node communication bandwidth is insufficient, decreasing PP reduces communication overhead, allowing more computation to be completed within the node.

[0056] `clip-grad`: Used to prevent gradient anomalies during training. Decrease this value to stabilize training if gradient explosion or NaN loss occurs; otherwise, use the default value when training is stable.

[0057] `train-iters`: This setting determines the total number of iterations for model training. Decrease this value to shorten the process when quickly verifying transfer learning; increase it to ensure sufficient convergence when fine-tuning the model's depth.

[0058] When users need to adjust parameter configurations, after the parameter recommendation agent in the intelligent decision layer receives the parameter configuration adjustment information, the agent analyzes and reasons based on multi-source information such as the basic parameter set, historical failure execution experience associated with the model migration task, computing power information of the target migration hardware, and decision guidance prompts. It then performs targeted optimization of the basic parameter set, such as adjusting core parameter configurations like tensor parallelism (TP), pipeline parallelism (PP), and layer allocation list (num_layer_list) as needed. Finally, it outputs an updated target migration parameter configuration scheme that adapts to the architectural features of the model to be migrated, the requirements of the model migration task, and the target migration hardware environment.

[0059] This fully leverages the logical reasoning and scenario understanding capabilities of large language models, and intelligently adjusts the basic parameter set based on the specific target migration computing power environment and historical migration execution experience to generate a target migration parameter configuration scheme that better meets actual needs.

[0060] After receiving confirmation from the user, the current model migration task is executed according to the target migration parameter configuration scheme or the updated target migration parameter configuration scheme.

[0061] Specifically, the current model migration process is pre-divided into several steps: weight transformation, model training, accuracy acquisition, and performance acquisition. Each step contains multiple sub-steps, and some sub-steps require the use of migration scripts in the target migration parameter configuration scheme to complete, such as the data preprocessing process and the model training process.

[0062] The steps requiring transfer scripts for weight transformation include model format conversion scripts (such as HF→MG, MG→HF), forward verification scripts for the target transfer hardware (such as NPU forward verification), and model training scripts for all sub-steps of model training.

[0063] Optionally, the target migration parameter configuration scheme can also be displayed to the user in the interactive access layer step by step, with each step executed sequentially to guide the user to complete the current model migration task step by step. This design simplifies the user operation to confirm the parameters in the migration script required for weight transformation and model training step by step and click to execute, without having to manually write migration scripts or operate in the command line. At the same time, it avoids the problems of insufficient scalability and high maintenance costs caused by traditional front-end hard-coded parameter logic, enabling the system to flexibly adapt to the differentiated configuration requirements of different models to be migrated in each step.

[0064] During the model migration task execution, the service coordination layer uses SSE long connections to push the execution logs and task progress of the computation execution layer to the interaction access layer in real time, enabling the monitoring of the migration status and the execution process of the model migration task to be presented synchronously. This design decouples business logic and computing resources into two independent layers: the service coordination layer handles all business processing, while the computation execution layer focuses on computation tasks. This allows the system to flexibly expand computing nodes without modifying core services, ensuring the versatility and maintainability of the system architecture.

[0065] Optionally, in one embodiment, the computation execution layer includes: a weight transformation module, a model training module, an accuracy acquisition module, and a performance acquisition module; executing the current model transfer task according to the target transfer parameter configuration scheme or the updated target transfer parameter configuration scheme includes: The weight conversion module is used to sequentially perform format conversion, weight consistency verification, original migration hardware forward verification, target migration hardware forward verification, and forward result comparison on the model to be migrated according to the model format conversion script and the forward verification script of the target migration hardware. The model training module is used to perform data preprocessing and pre-training iterations on the model to be transferred, according to the model training script. The accuracy acquisition module is used to provide the user with training process accuracy data when the user enables the training data acquisition function before the model to be transferred is pre-trained. The performance acquisition module is used to provide users with training process performance data when the user enables the training data acquisition function before the model to be transferred is pre-trained.

[0066] In this embodiment, the computation execution layer is deployed on the target migration hardware as an independent computing node. It receives the target migration parameter configuration scheme corresponding to the model migration task forwarded by the service coordination layer and completes the actual execution. This computation execution layer breaks down the complete process of the model migration task into four task modules: weight transformation, model training, accuracy acquisition, and performance acquisition.

[0067] Taking the original migration hardware as a GPU and the target migration hardware as an NPU as an example: The weight conversion module is responsible for format conversion (HF to MG format conversion, MG to HF format conversion), weight consistency verification, original migration hardware forward verification (GPU forward verification), target migration hardware forward verification (NPU forward verification), and forward result comparison.

[0068] The model training module completes data preprocessing and pre-training iterations, and supports users to enable the collection of training process accuracy data and training process performance data in the training data as needed before pre-training. After training is complete, users can go to the accuracy acquisition module and performance acquisition module to view the complete training process data, and check and optimize the inference performance of the model on the target transfer hardware (NPU) based on the training results.

[0069] Optionally, in one embodiment, the interactive access layer is also used to provide a custom script configuration entry point for models that have not yet been adapted; The custom script configuration entry is used to allow users to write and submit migration scripts for models that have not yet been adapted, and to store the migration scripts for models that have not yet been adapted in the script experience knowledge base for retrieval and reuse by subsequent model migration tasks of the same type.

[0070] For models not yet adapted to the system, the interactive access layer provides a custom script configuration entry point. Users can write and submit the parameters of the corresponding migration scripts according to the system's agreed format on each step page of the model migration task in the interactive access layer. The parameters of the submitted migration scripts will be automatically included in the migration script experience knowledge base for retrieval and reuse in subsequent similar model migration tasks, so that the system's model adaptation capability can be gradually expanded with use.

[0071] Based on the same inventive concept, another embodiment of this application also provides a large-model intelligent transfer method. Figure 4 This is a flowchart illustrating a large-scale intelligent transfer system method according to an embodiment of this application. (Refer to...) Figure 4 The method includes steps S11 to S16: Step S11: Store the historical execution experience and migration scripts of multiple historical model migration tasks in the script experience knowledge base; the historical execution experience includes historical successful execution experience and historical failed execution experience. Step S12: When the current model migration task is initiated, the script tool is used to search and determine whether there is any target historical successful execution experience that completely matches the current model migration task in the script experience knowledge base, and the search results are obtained: Step S13: If the search results indicate the existence of successful historical execution experience for the target, generate a target migration parameter configuration scheme for the current model migration task according to the successful historical execution experience for the target. Step S14: If the search results indicate that there is no successful historical execution experience of the target, the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, are fused with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme. Step S15: Use monitoring tools to collect the execution status, log information, and execution effect of the current model migration task. After migrating the model to be migrated to the target migration hardware according to the target migration parameter configuration scheme, generate optimization suggestions based on the execution status, log information, and execution effect of the current model migration task. Step S16: Using the monitoring tool, the target migration parameter configuration scheme and the optimization suggestions are used to generate new historical execution experience and store it in the script experience knowledge base.

[0072] Optionally, step S14, "integrating the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware, to generate a target migration parameter configuration scheme," specifically includes: Step S14-1: Receive the script tool generating search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated, and select candidate migration scripts from the script experience knowledge base based on the search keywords. Step S14-2: Receive the historical failure migration experience determined by the memory tool from the script experience knowledge base based on the search keywords; Step S14-3: Construct a basic parameter set based on the parameter names and default values ​​extracted from the candidate migration script; Step S14-4: Generate a target migration parameter configuration scheme based on the historical failed migration experience, the basic parameter set, the computing power information of the target migration hardware, and the architectural characteristics of the model to be migrated.

[0073] Optionally, step S14-1, "the script tool generates search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated, and determines candidate migration scripts from the script experience knowledge base based on the search keywords," specifically includes: Step S14-1-1: Obtain the configuration file of the model to be migrated and parse it to obtain the architectural features of the model to be migrated; Step S14-1-2: Based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated, generate search keywords, and based on the search keywords, initially screen out multiple initial candidate migration scripts that satisfy the search keywords from the script experience knowledge base: Step S14-1-3: Based on the configuration file of the model to be migrated, determine the first model series information and the first model size information of the model to be migrated, and based on the unique identifier of the model to be migrated, determine the second model series information and the second model size information of the model to be migrated. Step S14-1-4: Compare the first model series information and the first model size information with the second model series information and the second model size information to obtain the script filtering range, and filter the multiple initial candidate migration scripts based on the script filtering range to obtain multiple intermediate candidate migration scripts; Step S14-1-5: Using the first major language model, based on the computing power information of the target migration hardware, the functional type of the model to be migrated, and the configuration file of the model to be migrated, a candidate migration script is determined from the multiple intermediate candidate migration scripts.

[0074] Optionally, the step S14-2, "determining historical failure migration experiences from the script experience knowledge base based on the search keywords," specifically includes: Step S14-2-1: Generate an original query based on the search keywords, the migration steps, and the configuration information, and semantically expand the original query using the second language model to obtain an expanded query; Step S14-2-2: Based on the original query and the extended query, determine multiple initial candidate failed migration experiences from the script experience knowledge base; Step S14-2-3: Using a weighted fusion ranking method, based on the vector semantic similarity and time decay factor between the current model transfer task and the initial candidate failed transfer experiences, the top K initial candidate failed transfer experiences are determined as historical failed transfer experiences.

[0075] Optionally, the current model transfer task includes: weight transformation, model training, accuracy acquisition, and performance acquisition; the method further includes steps S21 to S26: Step S21: Receive the unique identifier of the model to be migrated from the user input. Step S22: Download the model to be migrated based on the unique identifier of the model; Step S23: Based on the model to be migrated downloaded from the service collaboration layer, obtain the architectural features of the model to be migrated and the migration steps of the model to be migrated; Step S24: Display the target migration parameter configuration scheme generated by the intelligent decision layer to the user, and receive the user's confirmation information or parameter configuration adjustment information on the target migration parameter configuration scheme in the interactive access layer; Step S25: After receiving the parameter configuration adjustment information, combine the parameter configuration adjustment information, the basic parameter set, the historical failed execution experience, the computing power information of the target migration hardware, the architectural features of the model to be migrated, and the preset decision guidance prompts. Optimize the basic parameter set through the third language model to obtain an updated target migration parameter configuration scheme and display it to the user through the interactive access layer until the user confirms the updated target migration parameter configuration scheme through the interactive access layer. Step S26: After receiving the confirmation information, execute the current model migration task according to the target migration parameter configuration scheme or the updated target migration parameter configuration scheme; The migration script in the target migration parameter configuration scheme or the migration script in the updated target migration parameter configuration scheme includes the model format conversion script required for weight conversion, the forward verification script for the target migration hardware, and the model training script required for model training.

[0076] Optionally, the computation execution layer includes: a weight transformation module, a model training module, an accuracy acquisition module, and a performance acquisition module; the step S26 of "executing the current model migration task according to the target migration parameter configuration scheme or the updated target migration parameter configuration scheme" specifically includes: Step S26-1: Perform format conversion, weight consistency verification, original migration hardware forward verification, target migration hardware forward verification, and forward result comparison on the model to be migrated in sequence according to the model format conversion script and the forward verification script of the target migration hardware. Step S26-2: According to the model training script, perform data preprocessing and pre-training iterations for the model to be transferred. Step S26-3: If the user enables the training data collection function before the model to be transferred is pre-trained, provide the user with training process accuracy data. Step S26-4: If the user enables the training data collection function before the model to be transferred is pre-trained, provide the user with training process performance data.

[0077] This application aims to verify the core performance and functional effectiveness of the parameter recommendation agent in a large-model intelligent transfer system during cross-hardware migration of a large language model from one primary transfer hardware to another, focusing on two dimensions: parameter configuration accuracy and weight transformation efficiency. Specific verification objectives are as follows: First, to evaluate the actual effect of the multi-source information fusion parameter recommendation method. The focus is on examining the mechanism's improvement in script selection accuracy and the rationality of parallel strategy recommendation, clarifying its actual contribution to optimizing decision results. Second, to quantitatively compare the efficiency differences between automated configuration and traditional manual configuration in the weight transformation stage. This verifies the advantage of the agent-based transfer system in end-to-end time consumption. Third, to verify the lightweight adaptability of the multi-source information fusion parameter recommendation method. Specifically, whether the mechanism, supported by a lightweight base model, can maintain stable, high-quality parameter recommendations through iterative optimization, thereby reducing dependence on the scale of base model parameters.

[0078] Based on the above objectives, the experimental scenario is designed as follows: 1) Models to be transferred: Three typical large language models based on LLaMA3-8B, Gemma-7B, and Qwen2.5-7B were selected as experimental subjects. The selected models have different architectures and a wide range of ecosystems, which can better verify the generality of the experimental results; 2) Hardware environment: NPU servers (e.g., equipped with Ascend 910B chips and Kunpeng 920 processors) are used to provide standardized hardware support for model transfer tasks; 3) Model transfer task: Weight conversion is the core evaluation task, and the end-to-end execution time is recorded. The model training stage is used to verify the accuracy of the transferred model and is not included in the time statistics. Only the rationality of the parameter recommendation script and hardware compatibility of this stage are evaluated.

[0079] 4) Baseline Method: Twenty professionals with experience in large language model migration and Ascend platform practice were selected, and a traditional manual configuration method was used as the baseline for comparison. The baseline method was based on a common initial script, and the professionals manually completed the entire process of script selection, parameter debugging, and weight conversion. The average time taken for the complete end-to-end process was calculated and used for quantitative comparison with the effect of the system's automated configuration. 5) Agent Base Model Setup: To explore the impact of base models with different parameter levels on agent recommendation performance, four typical parameter levels, Qwen3-0.6B, Qwen3-8B, Qwen3-32B, and Qwen3-80B, were selected as the agent's inference engine.

[0080] Experimental dataset: To fully verify the functionality and performance of NPU-Migrate-Agent, this application constructs a standardized script library and parameter recommendation test set to support the quantitative evaluation of parameter recommendation and migration efficiency.

[0081] 1) Standardized script library: as shown in Table 1, all of which are manually written and verified in real-world scenarios, providing reliable prior knowledge support for the parameter recommendation agent's migration script parameter recommendation and decision-making process; 2) Parameter Recommendation Test Set: A full-factor combination strategy of "model to be transferred × task type × computing power environment" was adopted to construct 240 sets of comprehensive migration request test cases. The test cases cover three typical models: LLaMA3-8B, Gemma-7B, and Qwen2.5-7B; two core tasks: weight transformation and pre-training verification; and four deployment configurations of the Ascend 910B chip: 1 card, 2 cards, 4 cards, and 8 cards per machine, ensuring the comprehensiveness and scenario coverage of the parameter recommendation effect evaluation. Meanwhile, to verify the lightweight adaptability of the multi-source information fusion parameter recommendation method, the above 240 sets of test cases were executed under four different intelligent agent base models with varying parameter levels, and all test cases were kept consistent.

[0082] Evaluation indicators: To achieve the experimental objectives, three core quantitative indicators were selected, and each indicator is defined as follows: 1) Average accuracy of transfer script selection: The proportion of transfer scripts output by the agent that match the characteristics of the model to be transferred and the hardware environment out of the total number of test samples. This metric is used to quantitatively evaluate the effectiveness of parameter recommendation methods based on multi-source information fusion in script adaptation decision-making; 2) Average Reasonableness of Parallel Strategy Recommendations: For three core parallel configuration parameters—Tensor Parallelism (TP), Pipeline Parallelism (PP), and layer allocation list (num_layer_list)—the percentage of recommended samples that simultaneously meet hardware resource constraints and model stability requirements after actual execution verification is statistically analyzed. This indicator is further subdivided into initial recommendation reasonableness and iterative recommendation reasonableness, which respectively measure the initial decision quality and iterative optimization capability of parameter recommendation methods based on multi-source information fusion. 3) Average total time for weight conversion: The average time taken from when the user initiates the model migration task to when the system completes automatic parameter generation and weight format conversion. The time taken for automated system configuration is compared with the average time taken for manual configuration by 20 professionals to quantitatively evaluate the improvement in migration efficiency.

[0083] Table 2: Average Accuracy of Transfer Script Selection and Core Parameter Configuration under Base Model Script matching accuracy 61.2% 98% 98% 100% PP / TP initial recommendation reasonableness 18% 36% 86% 98% Reasonableness of PP / TP Iterative Recommendations 90% 98% 100% 100% Layer allocation list matching accuracy 100% 100% 100% 100% Table 3: Comparison of average time for weight transformation of different models LLaMA3-8B 23min 9min Gemma-7B 26min 12min Qwen2.5-7B 25min 10min Experimental Results and Analysis: 1) Evaluation of the Support Effect and Lightweight Adaptability of the Parameter Recommendation Method Based on Multi-Source Information Fusion: Table 2 shows that the accuracy of transfer script matching and the rationality of the initial recommendation of PP / TP both increase with the increase of the number of base model parameters, reaching 100% and 98% respectively on Qwen3-80B. The accuracy of layer allocation list matching remains stable under all models, indicating that increasing the scale of the base model can significantly improve the initial recommendation quality of the parameter recommendation method based on multi-source information fusion, and this mechanism is robust in adapting to structured parameters such as layer allocation lists. The rationality of PP / TP iterative recommendation has reached 90% on Qwen3-0.6B, and further optimized with the increase of model scale, indicating that even on a lightweight base model, iterative optimization can still stably output high-quality parallel parameter configurations. The above results verify that the parameter recommendation method based on multi-source information fusion benefits from the increase of the base model scale on the one hand, and has good lightweight adaptability on the other hand, and is also effective under the support of conventional performance models.

[0084] 2) Weight Conversion Efficiency Comparison: Experimental results show that the weight conversion time of this system on all three test models is significantly lower than that of manual configuration. The average time is reduced from about 25 minutes for manual configuration to about 10 minutes, an efficiency improvement of about 60%. Moreover, the time fluctuation between models is significantly smaller than that of manual configuration. This indicates that this application can effectively shorten the model migration time and improve the overall migration efficiency and execution stability while ensuring migration correctness and hardware compatibility.

[0085] It should be noted that the large model intelligent migration method provided in this application embodiment can be executed by a large model intelligent migration device, or a control module in the large model intelligent migration device for executing the large model intelligent migration method. This application embodiment uses the large model intelligent migration device executing the large model intelligent migration method as an example to illustrate the large model intelligent migration method provided in this application embodiment.

[0086] The large-scale intelligent migration device in this application embodiment can be a device, or it can be a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network attached storage (NAS), personal computers (PCs), televisions (TVs), ATMs, or self-service machines, etc. The embodiments of this application do not impose specific limitations.

[0087] The large-model intelligent migration device in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0088] The large-model intelligent migration device provided in this application embodiment can achieve... Figure 4 The various processes implemented by the large model intelligent transfer device in the method embodiment will not be described again here to avoid repetition.

[0089] Optionally, Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. This application also provides an electronic device; it should be noted that the electronic device in this application includes the mobile electronic device and non-mobile electronic device described above.

[0090] The electronic device includes, but is not limited to, components such as: radio frequency unit, network module, audio output unit, input unit, sensor, display unit, user input unit, interface unit, memory, and processor.

[0091] Those skilled in the art will understand that electronic devices may also include power supplies (such as batteries) that supply power to various components. The power supply may be connected to the processor logic through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 5 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0092] As an example, such as Figure 5 As shown, the electronic device 600 includes a memory 610 and a processor 620. The memory 610 and the processor 620 are connected via a bus for communication. The memory 610 stores a computer program that can run on the processor 620 to implement the steps in the large model intelligent migration method disclosed in the above embodiments of this application.

[0093] As the apparatus is basically similar to the method embodiment, it is described in a relatively simple way. For relevant details, please refer to the description of the method embodiment.

[0094] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0095] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects.

[0096] Furthermore, this application embodiment also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described large model intelligent migration method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0097] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0098] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described large model intelligent migration method embodiment and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0099] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0100] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0101] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0102] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A large-scale intelligent transfer system, characterized in that, The system includes at least: A script experience knowledge base is used to store the historical execution experience and migration scripts of multiple historical model migration tasks; the historical execution experience includes historical successful execution experience and historical failed execution experience. The intelligent decision-making layer, when a current model migration task is initiated, uses script tools to search and determine whether there is any historical successful execution experience that perfectly matches the current model migration task in the script experience knowledge base, and obtains the search results: The intelligent decision-making layer is used to generate a target migration parameter configuration scheme for the current model migration task according to the target's historical successful execution experience when the search results indicate that there is such experience. The intelligent decision layer is used to, when the search results indicate that there is no successful historical execution experience of the target, integrate the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware, to generate a target migration parameter configuration scheme. The intelligent decision-making layer is used to collect the execution status, log information, and execution effect of the current model migration task using monitoring tools, and after migrating the model to be migrated to the target migration hardware according to the target migration parameter configuration scheme, it generates optimization suggestions based on the execution status, log information, and execution effect of the current model migration task. The intelligent decision-making layer is used to utilize the monitoring tool to generate new historical execution experience from the target migration parameter configuration scheme and the optimization suggestions and store it in the script experience knowledge base.

2. The large-scale intelligent transfer system according to claim 1, characterized in that, The step of fusing candidate migration scripts retrieved by the script tool from the script experience knowledge base, historical failed migration experiences associated with the current model migration task retrieved by the memory tool from the script experience knowledge base, with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme includes: The intelligent decision layer is used to receive the script tool generating search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model unique identifier of the model to be migrated, and to determine the candidate migration scripts from the script experience knowledge base based on the search keywords. The intelligent decision-making layer is used to receive historical failure migration experiences determined by the memory tool from the script experience knowledge base based on the search keywords; The intelligent decision-making layer is used to construct a basic parameter set based on the parameter names and default values ​​extracted from the candidate migration script; The intelligent decision-making layer is used to generate a target migration parameter configuration scheme based on the historical failed migration experience, the basic parameter set, the computing power information of the target migration hardware, and the architectural characteristics of the model to be migrated.

3. The large-scale intelligent transfer system according to claim 2, characterized in that, The script tool generates search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model's unique identifier. Based on these search keywords, it identifies candidate migration scripts from the script experience knowledge base, including: A script tool is used to obtain the configuration file of the model to be migrated and parse it to obtain the architectural features of the model to be migrated; A scripting tool is used to generate search keywords based on the architectural features of the model to be migrated, the migration steps of the model to be migrated, and the model's unique identifier. Based on these search keywords, it then initially filters out multiple candidate migration scripts that satisfy the search keywords from the script experience knowledge base. A script tool is used to determine the first model series information and the first model size information of the model to be migrated based on the configuration file of the model to be migrated, and to determine the second model series information and the second model size information of the model to be migrated based on the unique identifier of the model to be migrated. A script tool is used to compare the first model series information and the first model size information with the second model series information and the second model size information to obtain a script filtering range, and to filter the multiple initial candidate migration scripts based on the script filtering range to obtain multiple intermediate candidate migration scripts. A scripting tool is used to determine candidate migration scripts from a plurality of intermediate candidate migration scripts by utilizing the first major language model, based on the computing power information of the target migration hardware, the functional type of the model to be migrated, and the configuration file of the model to be migrated.

4. The large-scale intelligent transfer system according to claim 2, characterized in that, The historical failed migration experiences determined from the script experience knowledge base based on the search keywords include: An original query is generated based on the search keywords, the migration steps, and the configuration information. The original query is then semantically expanded using the second major language model to obtain an expanded query. Based on the original query and the extended query, multiple initial candidate failed migration experiences are determined from the script experience knowledge base; By using a weighted fusion ranking method, based on the vector semantic similarity and time decay factor between the current model transfer task and the initial candidate failed transfer experiences, the top K initial candidate failed transfer experiences are determined as historical failed transfer experiences.

5. The large-scale intelligent transfer system according to claim 4, characterized in that, The current model transfer task includes: weight transformation, model training, accuracy acquisition, and performance acquisition; The system also includes: The interactive access layer is used to receive the unique identifier of the model to be migrated from user input. A service collaboration layer is used to download the model to be migrated based on the model's unique identifier. The intelligent decision-making layer is used to obtain the architectural features of the model to be migrated and the migration steps of the model to be migrated based on the model to be migrated downloaded by the service collaboration layer. The interactive access layer is used to display the target migration parameter configuration scheme generated by the intelligent decision layer to the user, and to receive the user's confirmation information or parameter configuration adjustment information on the target migration parameter configuration scheme in the interactive access layer. The intelligent decision-making layer, upon receiving the parameter configuration adjustment information, combines the parameter configuration adjustment information, the basic parameter set, the historical failure execution experience, the computing power information of the target migration hardware, the architectural features of the model to be migrated, and preset decision guidance prompts. It then optimizes the basic parameter set through a third language model to obtain an updated target migration parameter configuration scheme, which is displayed to the user through the interactive access layer until the user confirms the updated target migration parameter configuration scheme through the interactive access layer. The computation execution layer is used to execute the current model migration task according to the target migration parameter configuration scheme or the updated target migration parameter configuration scheme after receiving the confirmation information; The migration script in the target migration parameter configuration scheme or the migration script in the updated target migration parameter configuration scheme includes the model format conversion script required for weight conversion, the forward verification script for the target migration hardware, and the model training script required for model training.

6. The large-scale intelligent transfer system according to claim 5, characterized in that, The computation execution layer includes: a weight transformation module, a model training module, an accuracy acquisition module, and a performance acquisition module; executing the current model transfer task according to the target transfer parameter configuration scheme or the updated target transfer parameter configuration scheme includes: The weight conversion module is used to sequentially perform format conversion, weight consistency verification, original migration hardware forward verification, target migration hardware forward verification, and forward result comparison on the model to be migrated according to the model format conversion script and the forward verification script of the target migration hardware. The model training module is used to perform data preprocessing and pre-training iterations on the model to be transferred, according to the model training script. The accuracy acquisition module is used to provide the user with training process accuracy data when the user enables the training data acquisition function before the model to be transferred is pre-trained. The performance acquisition module is used to provide training process performance data to the user when the user enables the training data acquisition function before the model to be transferred is pre-trained.

7. The large-scale intelligent transfer system according to claim 5, characterized in that, The interactive access layer is also used to provide a custom script configuration entry point for models that have not yet been adapted; The custom script configuration entry is used to allow users to write and submit migration scripts for models that have not yet been adapted, and to store the migration scripts for models that have not yet been adapted in the script experience knowledge base for retrieval and reuse by subsequent model migration tasks of the same type.

8. A method for intelligent transfer learning of large models, characterized in that, The method includes: The historical execution experience and migration scripts of multiple historical model migration tasks are stored in the script experience knowledge base; the historical execution experience includes historical successful execution experience and historical failed execution experience. When the current model migration task is initiated, the script tool is used to search and determine whether there is any historical successful execution experience that completely matches the current model migration task in the script experience knowledge base, and the search results are obtained: If the search results indicate the existence of successful historical execution experience for the target, a target migration parameter configuration scheme for the current model migration task is generated according to the successful historical execution experience for the target. If the search results indicate that there is no successful historical execution experience of the target, the candidate migration scripts retrieved by the script tool in the script experience knowledge base, the historical failed migration experiences associated with the current model migration task retrieved by the memory tool in the script experience knowledge base, are fused with the architectural features of the model to be migrated corresponding to the current model migration task and the computing power conditions of the target migration hardware to generate a target migration parameter configuration scheme. The execution status, log information, and execution effect of the current model migration task are collected using monitoring tools. After the model to be migrated is migrated to the target migration hardware according to the target migration parameter configuration scheme, optimization suggestions are generated based on the execution status, log information, and execution effect of the current model migration task. Using the monitoring tool, the target migration parameter configuration scheme and the optimization suggestions are used to generate new historical execution experience and stored in the script experience knowledge base.

9. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps in the large model intelligent migration method as described in claim 8.

10. A readable storage medium, characterized in that, The program or instructions are stored on the readable storage medium, and when the program or instructions are executed by the processor, they implement the steps in the large model intelligent migration method as described in claim 8.

Images