Big model federal fine tuning and reasoning collaborative execution method oriented to edge intelligence

Abstract

The invention discloses an edge intelligence-oriented large model federated fine tuning and reasoning cooperative execution method, and aims to solve the problems of resource waste and response delay caused by separation of fine tuning and reasoning of a large model in an edge intelligence scene. According to the method, on the basis of a federated learning framework, an incremental weight is dynamically calculated by keeping an LoRA adapter independent, a movable adapter is used for reasoning through a mirror image adapter strategy, a mirror image adapter is used for fine tuning, then parameters are synchronized through atomic exchange, and finally cooperative proceeding of fine tuning and reasoning in a single copy is achieved; meanwhile, a double-time-scale coordination algorithm is put forward to optimize and finely adjust the batch size and the reasoning rate on a long time scale, reasoning requests are dynamically scheduled on a short time scale, and finally a multi-copy task is coordinated. According to the method, repeated deployment is eliminated, resource occupation is reduced, the inference service obtains fine adjustment benefits in real time, long and short term targets are balanced, the inference quality and throughput are remarkably improved, and the method is suitable for edge deployment of various large models.

Description

Technical Field

[0001] This invention relates to the field of large-scale artificial intelligence model post-training technology, specifically a method for federated fine-tuning and inference collaborative execution of large-scale models for edge intelligence. Background Technology

[0002] With the development of artificial intelligence technology, large-scale language models, large-scale computer vision models, and large-scale speech processing models have been widely used in fields such as coding assistants, image generation, and voice assistants. The performance of these models depends on the optimization of the entire process of "pre-training-fine-tuning-inference": pre-training builds the basic capabilities of the model, while fine-tuning and inference, as the post-training stage, directly determine the quality of service provided by the model to end users.

[0003] To protect data privacy, avoid uploading local data from edge nodes to the cloud, and adapt to the diversity of edge data distribution, accent differences in voice data, and domain differences in text data, existing technologies often adopt a solution that combines LoRA parameter fine-tuning technology with a federated learning framework for edge deployment of various large-scale models.

[0004] However, existing technologies generally treat fine-tuning and inference as separate workloads. That is, the model is first fine-tuned at the edge or in the cloud, and then the fine-tuned model is redeployed to the edge nodes for inference services. Although this separate processing approach can achieve model updates, it has the following common drawbacks in resource-constrained edge environments, such as when edge servers have limited memory and GPU computing power: Redundant resource consumption: Fine-tuning and inference processes each require loading the complete model backbone and LoRA adapter, repeatedly occupying the resources of edge nodes, resulting in low resource utilization.

[0005] Model update delay: The inference service must wait for fine-tuning to be completed and the model to be redeployed before it can use the latest fine-tuning parameters. This serial process of "fine-tuning-deployment-inference" causes users to be unable to obtain the optimized model service in a timely manner.

[0006] Lack of resource scheduling mechanism: Traditional system scheduling strategies mainly focus on efficiency indicators such as throughput, while ignoring the improvement in inference service quality brought about by fine-tuning. Under limited resource conditions, the system is unable to make an effective trade-off and joint scheduling between the short-term goal of responding to user requests in a timely manner and the long-term goal of continuously improving model quality through fine-tuning, thus failing to maximize the overall service efficiency.

[0007] In summary, regardless of whether it is a large language model, computer vision model, or speech model, the separate processing of fine-tuning and inference in edge environments will lead to resource waste, response latency, and ineffective resource scheduling. Therefore, there is an urgent need for a federated fine-tuning and inference collaborative execution method for large models in edge intelligence, which can realize unified collaboration of fine-tuning and inference of various large models in resource-constrained edge clusters, and make full use of the quality improvement brought by fine-tuning in real time while ensuring low latency. At the same time, it can coordinate tasks of multiple edge replicas to achieve a balance between short-term and long-term goals. Summary of the Invention

[0008] To address the problems existing in the background technology, this invention proposes a large-model federated fine-tuning and inference collaborative execution method for edge intelligence. This method solves the problems of redundant resource consumption, model update delay, and ineffective resource scheduling caused by the separation of fine-tuning and inference in edge intelligence scenarios. It achieves unified collaboration of large-model fine-tuning and inference in edge clusters, resulting in low resource consumption, low latency, and high-quality real-time services.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A collaborative execution method for federated fine-tuning and inference of a large model for edge intelligence, applied to an edge cluster including a central server and several edge replicas, includes the following steps: S1. The central server deploys the global large model, and each edge replica loads the frozen global large model backbone and configures a LoRA adapter with efficient parameter fine-tuning. S2. Based on the predicted request arrival rate and the resource status of each edge replica, the central server separates short-term inference scheduling from long-term fine-tuning decisions using a dual-time-scale coordination algorithm, and determines the planned inference rate and fine-tuning batch size for each edge replica. S3. Each edge copy uses the Unmerged inference mechanism and the mirror adapter strategy to uniformly and collaboratively execute the model fine-tuning and inference process, and adjust the LoRA adapter parameters in real time. S4. Each edge replica periodically uploads the LoRA adapter parameters to the central server. The central server performs weighted aggregation based on the local data volume of each edge replica to obtain the global LoRA adapter parameters, updates the global large model, and distributes it to each edge replica to update its LoRA adapter parameters. S5. Repeat steps S2 to S4 to fine-tune the global large model.

[0010] Specifically, the LoRA adapter in step S1 includes a low-rank matrix. and low-rank matrix The dimension is The dimension of the low-rank matrix A is ,in For rank, and These are the input and output dimensions of the global large model, respectively, calculated... Generate incremental weights.

[0011] Specifically, the dual-timescale coordination algorithm in step S2 includes: (1) Optimization of the federated rounds: In each federal round Initially, the central server calculates based on the predicted request arrival rate. And the resource status of the edge replicas, for each edge replica Determine the planned inference rate and fine-tuning batch size Planning inference rate and fine-tuning batch size With the goal of balancing current inference output with future model quality improvement and maximizing quality-aware throughput, the objective function is: ; in, Let S be the set of edge copies, and S be the number of fine-tuning iterations within each federation round R. For edge copy The time consumed by a single round of fine-tuning For edge copy In the Rounds The quality of inference in each iteration. For edge copy In the The amount of local data for each round of fine-tuning. As a balancing factor, it is used to weigh the output of the current round of inference against the future improvement of model quality. Indicates edge copy Local fine-tuning improves the overall quality of the large model. Adjust batch size and number of iterations The evolution formula is: ; in For edge copy Inference quality adjustment parameters, The calculation formula is: ; in and It is with the federal round The relevant adjustment parameters, For edge copy In the Rounds The quality of inference in each iteration. For edge copy In the Rounds The quality of reasoning in each iteration; The constraints on the objective function include: the planned inference rate. The request arrival rate requirement is met, and the sum of the memory consumption for training and inference does not exceed the edge replica. memory capacity ; (2) Slot-level inference scheduling: In federal rounds Determine the planned inference rate and fine-tuning batch size Then, the central server in each time slot The system performs batch scheduling of actual arriving inference requests, with scheduling rules including: a. Form batches according to the first-in-first-out queue requests. Batch size satisfy: ; in, For edge copy Processing batches The execution time includes inference time and communication latency; b. Batch Execution time The deadline for all requests within the batch shall not exceed the deadline for all requests within the batch; c. When a service quality violation of latency targets is detected, or the actual request arrival rate deviates from the planned inference rate. When a mismatch occurs, a reactive correction mechanism is used to dynamically adjust the fine-tuning batch size. and planning reasoning rate .

[0012] Specifically, quality-perceived throughput is a metric that measures system throughput and inference quality, and its calculation formula is as follows: ; in Time interval The number of inference requests successfully completed within the time limit. For reasoning quality.

[0013] Specifically, the optimal planning inference rate is solved by the objective function. and fine-tuning batch size The specific method is as follows: Set a finite set of candidate fine-tuning batch sizes; within the finite set of candidate fine-tuning batch sizes, calculate the objective function value corresponding to each candidate fine-tuning batch size by enumeration; and select the planned inference rate corresponding to the maximum value. and fine-tuning batch size .

[0014] Specifically, the Unmerged inference mechanism in step S3 is as follows: Preserve the low-rank matrix of the LoRA adapter low-rank matrix Independent of the frozen global backbone weights, incremental weights are dynamically calculated during inference for each inference request's forward computation process. and will The weights are added to the frozen global backbone model to obtain the current model weights, and then inference is performed.

[0015] Specifically, the mirror adapter strategy in step S3 is as follows: Each edge replica maintains two LoRA adapter instances: an active adapter and a mirror adapter. The active adapter includes a low-rank matrix. The mirror adapter includes a low-rank matrix. and ; During inference, use the low-rank matrix of the active adapter. conduct Dynamic calculation, will Add the weights to the frozen global backbone model to obtain the current model weights, and then perform inference. When fine-tuning, use the mirror adapter. and Perform gradient calculations and update the parameters of the mirror adapter; After each fine-tuning step, the parameters of the active adapter and the mirror adapter are swapped through atomic operations, and the current parameters of the active adapter are synchronized to the mirror adapter, and then real-time inference and fine-tuning continue.

[0016] In summary, the beneficial technical effects of the present invention are as follows: 1. Eliminate redundant deployments and reduce resource consumption: Through the unmerged inference mechanism and the image adapter strategy, fine-tuning and inference share parameters on a single model copy, avoiding repeated loading of models and adapters, and significantly reducing memory consumption and model redeployment time.

[0017] 2. Real-time adaptation and improved inference quality: The mirror adapter immediately applies the latest parameters to inference through atomic swaps after fine-tuning, enabling the inference service to obtain the benefits of fine-tuning in real time and eliminating the delay of waiting for model redeployment in traditional methods.

[0018] 3. Quality-aware joint scheduling: By using a dual-time-scale coordination algorithm, short-term and long-term goals are balanced under limited resources, which ensures both the throughput of immediate inference and the quality of models in future rounds.

[0019] 4. Scalability and universality: The model sharing mechanism and coordination algorithm designed in this invention do not depend on a specific large model structure and can be directly applied to various federated fine-tuning tasks based on LoRA adapters. It is suitable for deploying a variety of large models in edge clusters. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention clearer and easier to understand, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0022] Example like Figure 1 As shown, the method for collaborative execution of large-scale model federated fine-tuning and inference for edge intelligence provided by this invention specifically includes the following steps: Step S1: The central server first deploys the pre-trained global large model. Each edge replica loads the frozen global large model, whose backbone parameters are fixed and do not participate in fine-tuning, thus maintaining the basic capabilities of the large model and avoiding the destruction of pre-trained knowledge during fine-tuning. At the same time, each edge replica is configured with a LoRA adapter for efficient parameter fine-tuning. LoRA adapters include low-rank matrices. and low-rank matrix The dimension is low-rank matrix The dimension is ,in For rank, and These are the input and output dimensions of the global large model, respectively, calculated... Incremental weights are generated and the model is adjusted using a low-rank matrix, with training performed only. and This enables efficient parameter fine-tuning and reduces the computational burden on edge replicas.

[0023] Step S2: Based on the predicted request arrival rate, i.e. the average request volume over a future period, and the resource status of each edge replica, such as memory and computing power, the central server separates short-term inference scheduling from long-term fine-tuning decisions using a dual-time-scale coordination algorithm. It determines the planned inference rate and fine-tuning batch size for each edge replica to maximize the system's quality-aware throughput and meet resource and latency constraints.

[0024] Assuming at time Edge copy There is a set of requests to be processed in the inference queue. Each request has a deadline. ,set up For a batch request, For batch size, To use batch size After minor adjustments in the instance upper processing batch To determine the desired inference quality, the joint scheduling can be modeled as a mixed integer nonlinear programming (MINLP) as follows: The constraints include (a) ensuring that each request is scheduled at most once, (b) ensuring that a batch is not executable before it arrives, (c) ensuring that each batch completes before the deadline, and (d) ensuring that GPU memory does not exceed its capacity. (e) Ensure that each copy processes only one batch at a time.

[0025] Since the MINLP is difficult to solve in an online environment, this invention further proposes a two-stage coordination algorithm, namely the dual-time-scale coordination algorithm, which separates long-term fine-tuning decision-making from short-term inference scheduling.

[0026] The dual-time-scale coordination algorithm is divided into federated round-level optimization and slot-level inference scheduling: (1) Optimization of the federated rounds: A federated round refers to the entire cycle from the start of one global model aggregation to the start of the next. It is the basic time unit for collaborative training in federated learning. Initially, the central server calculates based on the predicted request arrival rate. And the resource status of the edge replicas, for each edge replica Determine the planned inference rate and fine-tuning batch size Planning inference rate and fine-tuning batch size The goal is to balance current inference output with future model quality improvement, and to maximize quality-aware throughput. Quality-aware throughput is a metric that measures system throughput and inference quality, and its calculation formula is as follows: ; in Time interval The number of inference requests successfully completed within the time limit. For the quality of reasoning; Planned inference rate and fine-tuning batch size The objective function is: ; in, Let S be the set of edge copies, and S be the number of fine-tuning iterations within each federation round R. For edge copy The time consumed by a single round of fine-tuning For edge copy In the Rounds The quality of inference in each iteration. For edge copy In the The amount of local data for each round of fine-tuning. As a balancing factor, it is used to weigh the output of the current round of inference against the future improvement of model quality. Indicates edge copy Local fine-tuning improves the overall quality of the large model. Adjust batch size and number of iterations The evolution formula is: ; in For edge copy Inference quality adjustment parameters, The calculation formula is: ; in and It is with the federal round The relevant adjustment parameters, For edge copy In the Rounds The quality of inference in each iteration. For edge copy In the Rounds The quality of reasoning in each iteration; The constraints of the objective function include: planned inference rate. The request arrival rate requirement is met, and the sum of the memory consumption for training and inference does not exceed the edge replica. memory capacity ; The optimal planning inference rate is obtained by solving the objective function. and fine-tuning batch size The specific method is as follows: Set a finite set of candidate fine-tuning batch sizes; within the finite set of candidate fine-tuning batch sizes, calculate the objective function value corresponding to each candidate fine-tuning batch size by enumeration; and select the planned inference rate corresponding to the maximum value. and fine-tuning batch size This method can effectively reduce computational complexity and achieve efficient online solutions.

[0027] The federated round of secondary optimization can ultimately balance short-term and long-term goals through... We need to balance current inference output with future quality improvement, avoid extreme situations where we prioritize short-term gains over long-term gains or vice versa, and optimize resources by making adaptive adjustments based on the resources available for edge instances.

[0028] (2) Slot-level inference scheduling: In federal rounds Determine the planned inference rate and fine-tuning batch size Then, the central server in each time slot The system performs batch scheduling of actual arriving inference requests, with scheduling rules including: a. Batch assembly: Assemble batches according to the first-in-first-out queue requests. Batch size satisfy: ; in, For edge copy Processing batches The execution time includes inference time and communication latency; b. Deadline Guarantee: Batch Execution time The deadline for all requests within the batch shall not exceed the deadline for all requests within the batch; c. Reactive Correction: When a service quality violation of latency targets is detected, or the actual request arrival rate deviates from the planned inference rate... When a mismatch occurs, a reactive correction mechanism is used to dynamically adjust the fine-tuning batch size. and planning reasoning rate .

[0029] Ultimately, slot-level inference scheduling ensures that inference requests are completed within the deadline by adjusting the batch size. When the actual situation does not match the plan, resource allocation is adjusted in a timely manner, such as reducing the fine-tuned batch size to release resources for inference, thus enabling flexible short-term decision-making.

[0030] Furthermore, through the aforementioned dual-timescale algorithm, this invention optimizes the fine-tuning batch size and inference rate on a long-term timescale to improve long-term model quality, and dynamically schedules inference requests on a short-term timescale to ensure real-time service quality, thereby achieving collaborative optimization of fine-tuning and inference in edge clusters.

[0031] Step S3: Each edge copy uses the Unmerged inference mechanism and mirror adapter strategy to uniformly and collaboratively execute the model fine-tuning and inference process, and adjust the LoRA adapter parameters in real time. The Unmerged reasoning mechanism is as follows: Edge copies do not merge low-rank matrices of LoRA during inference. low-rank matrix Maintain the low-rank matrix of the LoRA adapter low-rank matrix Independent of the frozen global backbone weights, incremental weights are dynamically calculated for the forward computation process of each inference request. and will The weights are added to the weights of the frozen global backbone model to obtain the current model weights. Perform inference.

[0032] The specific mirror adapter strategy is as follows: Each edge replica maintains two LoRA adapter instances: an active adapter and a mirror adapter. The active adapter includes a low-rank matrix. The mirror adapter includes a low-rank matrix. and ; During inference, use the low-rank matrix of the active adapter. conduct Dynamic calculation, will Add the weights to the frozen global backbone model to obtain the current model weights, and then perform inference. When fine-tuning, use the mirror adapter. and Perform gradient calculations and update the parameters of the mirror adapter; After each fine-tuning step, the parameters of the active adapter and the mirror adapter are swapped through atomic operations, and the current parameters of the active adapter are synchronized to the mirror adapter, and then real-time inference and fine-tuning continue.

[0033] Furthermore, within a single edge copy, this invention achieves the sharing of global large model backbone parameters between fine-tuning and inference through unmerged inference and mirror adapter strategies, resulting in the following effects: Structural consistency: maintain and Independent and not merged, ensuring that inference can still be performed during fine-tuning; Parameter consistency: Read and write are isolated through mirror adapters to avoid data races. Inference only accesses stable active adapters. Fine-tuning performs gradient calculations on mirror adapters and switches after each step to ensure that inference always uses a complete set of parameters. Controllable overhead: The memory overhead of dual adapters is... , For adapter rank, The input and output dimensions account for only about 1% of the base model parameters.

[0034] Step S4: Periodic federated aggregation to update the global model: Each edge replica periodically uploads its LoRA adapter parameters to the central server. The central server performs weighted aggregation based on the local data volume of each edge replica to obtain the global LoRA adapter parameters, updates the global model, and distributes it to each edge replica to update its LoRA adapter parameters.

[0035] Step S5: Repeat steps S2 to S4 to fine-tune the global model. As time goes on, the local data of the edge replicas continues to increase. The cyclical fine-tuning enables the model to continuously adapt to new data. Finally, through multiple rounds of federated aggregation, the model gradually converges and the inference quality gradually improves.

[0036] Therefore, this invention provides a method for collaborative execution of federated fine-tuning and inference of large models for edge intelligence. This method employs an Unmerged inference mechanism to maintain the independence of LoRA adapters and model backbone parameters, a mirror adapter strategy to isolate access to fine-tuning and inference parameters, a dual-timescale coordination algorithm to separate long-term fine-tuning decisions from short-term inference scheduling, and finally updates the global model through federated aggregation. This achieves a closed-loop collaborative execution process of "parameter sharing - collaborative execution of inference fine-tuning - long-term and short-term balancing - global optimization," solving the pain points of redundant resource consumption, model update delays, and the inability to effectively balance long-term and short-term goals caused by the traditional separation of fine-tuning and inference. It improves the efficiency of edge resource utilization, the real-time performance of inference quality improvement, and the scientific nature of scheduling decisions. It significantly reduces memory consumption, model redeployment time, and the probability of conflicts between long-term and short-term goals, ensuring the continuity and low latency of model services in edge clusters. Finally, through a scalable model sharing mechanism and a general coordination algorithm, it provides an efficient and reliable solution for the collaborative federated fine-tuning and inference of various large models in edge intelligence scenarios, ensuring high-quality large model services in resource-constrained edge environments.

[0037] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An edge intelligence-oriented large model federated fine-tuning and inference collaborative execution method applied to an edge cluster comprising a center server and a plurality of edge replicas, characterized in that, The method comprises the following steps: S1, the center server deploys a global large model, each edge copy loads a frozen global large model backbone, and configures a LoRA adapter for efficient parameter fine-tuning; S2, the center server separates short-term inference scheduling from long-term fine-tuning decision according to the predicted request arrival rate and the resource condition of each edge copy through a double-time-scale coordination algorithm, determines the planned inference rate and fine-tuning batch size for each edge copy; S3, each edge copy unifies and cooperatively executes the model fine-tuning and inference process through an Unmerged inference mechanism and a mirror adapter strategy, and adjusts the LoRA adapter parameters in real time; S4, each edge copy uploads the LoRA adapter parameters to the center server at regular intervals, the center server aggregates the global LoRA adapter parameters according to the local data volume of each edge copy, updates the global large model, and then delivers the updated LoRA adapter parameters to each edge copy for updating; S5, repeating steps S2 to S4 to fine-tune the global large model.

2. The method of claim 1, wherein, The LoRA adapter in step S1 includes a low-rank matrix and The dimension of the low-rank matrix The dimension of the low-rank matrix wherein is the rank, and are the input dimension and the output dimension of the global large model respectively, and the incremental weight is generated by calculating .​​ 3. The method of claim 1, wherein, The double-time-scale coordination algorithm in step S2 comprises: (1) Federal round secondary optimization: In each federal round Initially, the central server calculates based on the predicted request arrival rate. And the resource status of the edge replicas, for each edge replica Determine the planned inference rate and fine-tuning batch size Planning inference rate and fine-tuning batch size With the goal of balancing current inference output with future model quality improvement and maximizing quality-aware throughput, the objective function is: ； wherein, is a set of edge replicas, S is the number of fine-tuning iterations within each federation round R, is an edge replica single round fine-tuning time consumption, is an edge replica inference quality of the edge replica in the first round the th iteration, is a local data volume of the edge replica fine-tuned in the first round, is a trade-off factor for balancing the weight of current round inference output and future model quality improvement, denotes the improvement of global large model quality by local fine-tuning of the edge replica , With evolution of the fine-tuning batch size and the number of iterations The evolution formula is: ； wherein is an edge copy of the inference quality adjustment parameter, The calculation formula is: ； wherein and are regulatory parameters related to the federal round , is the edge copy at the first round the first iteration of the inference quality, is the edge copy at the first round the first iteration of the inference quality; The constraints of the objective function include: planning inference rate satisfying the request arrival rate requirement, and the sum of memory consumption of training and inference does not exceed the memory capacity of the edge replica ;​ (2) Slot-level inference scheduling: At the federal round Determining a planned reasoning rate And fine-tuning batch size After that, the central server batches the actual arrived reasoning requests in each time slot And schedules them according to the following rules: a. Extract requests into batches according to a first-in-first-out queue , batch size satisfies: ； wherein, edge replica processing batch execution time, including inference latency and communication delay; b. Batch Time of execution No more than the deadline of all requests within a batch; c. When a service quality violation is detected, the delay target or actual request arrival rate is compared to the planned inference rate When not matched, a reactive correction mechanism is employed to dynamically adjust the fine-tuning batch size and the planned inference rate .

4. The method of claim 3, wherein, The quality-aware throughput is a quality-aware throughput, and the calculation formula is: ； wherein is a time interval number of inference requests successfully completed within is a quality of inference.

5. The method of claim 3, wherein the method is characterized in that, Solving the optimal planning inference rate through the objective function And fine-tuning batch size The method is specifically: setting a limited candidate fine-tuning batch size set, in the limited candidate fine-tuning batch size, calculating the objective function value corresponding to each candidate fine-tuning batch size through enumeration, and selecting the planning inference rate corresponding to the maximum value And fine-tuning batch size .

6. The method of claim 2, wherein the method is characterized in that, The Unmerged inference mechanism in step S3 is: Keeping the low rank matrix of LoRA adapter , low rank matrix Independent of the frozen global large model backbone weight, during inference, the forward calculation process for each inference request, dynamically calculate the incremental weight , and add to the weight of the frozen global large model backbone, get the current model weight, and execute inference.

7. The method of claim 6, wherein, The mirror adapter strategy in step S3 is: Each edge replica maintains two LoRA adapter instances, an active adapter comprising a low-rank matrix and a mirror adapter comprising a low-rank matrix and ​ Inference, using low rank matrix of active adapter Perform Dynamic calculation of Add to the weight of the frozen global large model backbone to get the current model weight, and perform inference; Fine-tuning, using the mirror adapter and Gradient computation and updating of the mirror adapter's parameters; After each fine-tuning step, the parameters of the active adapter and the mirror adapter are exchanged through an atomic operation, the current parameters of the active adapter are synchronized to the mirror adapter, and then real-time inference and fine-tuning are continued.

Images