A convergence-aware multi-tenant edge large model adaptive fine-tuning scheduling method and system

Abstract

The application discloses a convergent-aware multi-tenant edge large model adaptive fine-tuning scheduling method and system, taking a GPU resident adapter slot in a device of an edge node as a resource abstract object, under the constraints that a video memory can only accommodate a limited number of trainable adapters at the same time, and the computing power is only allowed to perform fine-tuning steps sequentially, online calculation allocation, convergence identification, degradation detection and state resetting are performed on personalized adapters of multiple tenants. By adopting the technical scheme of the application, under the premise of not relying on complex global joint optimization, the efficient allocation of limited training opportunities is realized by using a deployable system structure and an online scheduling mechanism.

Description

Technical Field

[0001] This invention belongs to the field of resource scheduling optimization technology, specifically relating to an adaptive fine-tuning scheduling method and system for a convergent-aware multi-tenant edge large model. Background Technology

[0002] With the increasing demand for deploying large language models on edge devices and near-end nodes, directly retaining user data at the edge and performing efficient parameter fine-tuning has become an important technical approach that balances privacy, latency, and personalization capabilities. Unlike a unified large model in the cloud, edge nodes often need to serve multiple tenants or multiple task domains simultaneously, and the data distribution, update frequency, task objectives, and available training sample size vary significantly among different tenants. To reduce training costs, existing systems typically employ efficient parameter fine-tuning methods, such as adapters, prefix / prompt tuning, and LoRA, updating only a few additional parameters while freezing the core shared base model.

[0003] However, multi-tenant edge fine-tuning differs significantly from traditional single-task fine-tuning, traditional inference caching, or general job scheduling. First, edge devices have severely limited GPU memory. The fine-tuning process requires storing not only adapter parameters but also training working sets such as gradients, optimizer states, and instantaneous activations. Therefore, typically only a small number of tenant adapters can be kept in a "directly trainable" state. Second, edge devices have limited computing power and power budgets, making it difficult to perform gradient updates on multiple tenants simultaneously in parallel. In practice, the training pipeline is often allocated to only one tenant at a time. Third, multi-tenant scenarios exhibit significant non-stationarity: a tenant adapter gradually converges after continuous training, and its marginal returns decrease; simultaneously, tenant replacement, task switching, user relay, or data distribution drift can quickly invalidate existing statistical information.

[0004] The existing technology has the following main shortcomings: (1) Many methods jointly model problems such as adapter admission / exit, training scheduling, unloading and resource allocation. Although such schemes have complete objectives, they have large state spaces and too many coupling factors, making it difficult to complete online solutions with low latency on edge devices, which is not conducive to high-frequency training decisions.

[0005] (2) Common random multi-armed gambling machine or general non-stationary gambling machine methods usually rely on passive forgetting mechanisms, such as fixed windows or discount coefficients. Such methods do not explicitly distinguish between the two cases of "the tenant itself has low value" and "the tenant has basically converged", and are prone to continuously wasting computation on adapters that have no obvious room for improvement.

[0006] (3) Some change detection methods can handle abrupt changes, but they do not combine the convergence law and adapter replacement mechanism in edge fine-tuning, and cannot simultaneously utilize the two structural information of "the benefits gradually decay until saturation" and "the slot becomes valuable again due to remapping".

[0007] (4) Existing studies generally ignore the state transition overhead caused by training context switching. For edge systems, frequent switching of training targets will cause the adapter and optimizer states to be repeatedly switched in and out between the GPU and CPU / SSD, resulting in additional I / O, energy consumption and stability loss.

[0008] Therefore, a specific technical solution is needed that is truly geared towards edge multi-tenant continuous fine-tuning scenarios, which can achieve efficient allocation of limited training opportunities with a deployable system architecture and online scheduling mechanism without relying on complex global joint optimization. Summary of the Invention

[0009] This invention provides a convergence-aware adaptive fine-tuning scheduling method and system for large multi-tenant edge models, addressing the following technical problems: 1) Given that the GPU memory can only accommodate a limited number of trainable tenant adapters and only one slot can be selected to perform fine-tuning steps in each round, how can the limited training opportunities be preferentially allocated to the tenant that is most worthy of training at the moment?

[0010] 2) As the training benefits of the tenant adapter gradually decrease with each training round and eventually reach saturation, how can we identify the moment when "continuing training is no longer worthwhile" to avoid invalid updates?

[0011] 3) When a tenant is swapped out or into a slot by the system layer, or when the data distribution of the same tenant changes abruptly, how can we quickly identify that old statistics are outdated and reset historical information in a timely manner to restore exploration capabilities?

[0012] 4) Without changing the existing edge-side adapter caching / swapping mechanism, how to provide a pluggable learning layer scheduler that can run based solely on current slot occupancy and online feedback.

[0013] 5) While ensuring scheduling feasibility, we also take into account model quality improvement, computing resource utilization and context switching overhead, so as to improve the overall quality and efficiency of edge personalized services.

[0014] To achieve the above objectives, the present invention provides the following solution: A convergence-aware adaptive fine-tuning scheduling method for large multi-tenant edge models includes: Step S1: System initialization, setting the number of slots Rewards Window Convergence detection window Degradation detection window Exploration coefficient Confidence parameters Single-step adaptation cost Convergence threshold and degradation threshold Initialize the counter for each slot. Status flags Historical tenant identifier And clear the reward window, earnings window and agent window; Step S2: In round Initially, the current slot occupancy mapping is read from the system layer. If the current tenant of a slot is found to be the same as the previous record... Inconsistency indicates that tenant remapping or slot refilling has occurred at the system level; in this case, updating is necessary. And immediately perform a degradation reset on the slot to clear the old history; Step S3: Calculate the scheduling index for all slots; for slots that have been marked as saturated, set their index to negative infinity; for slots that have no samples yet, set their index to positive infinity; for normal slots, calculate the index by summing the window mean and the exploration items. Step S4: Select the slot with the highest index. For the tenants Perform one incremental fine-tuning step; Step S5: Extract incremental gains from this training step With agent signal ,according to Calculate net rewards and update the recent rewards window, earnings window, and agent window; Step S6: If the profit window is full, check if the convergence / saturation conditions are met; if they are met, immediately perform a saturation reset on the selected slot and temporarily freeze it. Step S7: If the agent window is full, compare the average of its front and back segments to determine if a degradation event has occurred; if the degradation conditions are met, immediately perform a degradation reset on the selected slot. Step S8: Proceed to the next round, repeating the above process until the set number of rounds is reached or the system stops running. This invention also provides a convergence-aware multi-tenant edge large model adaptive fine-tuning scheduling system, comprising: The system-level admission / exit module is used to determine whether a tenant adapter enters or leaves a GPU slot based on cache pressure, service priority, and tenant arrival / departure factors on a slow time scale, and to form a mapping relationship between the current slot and the tenant. The learning layer scheduling module is used to select a slot for incremental fine-tuning in each fast round based on the most recent reward statistics and confidence upper bound of each slot; The benefit observation module is used to extract the incremental performance benefit brought by the fine-tuning step after a fine-tuning step is completed in the selected slot. ; The agent monitoring module is used to extract lightweight agent signals. ; The convergence perception detection module is used to perform a moving average of the recent incremental gains. When the moving average is lower than a set threshold, it is determined that the slot has converged or saturated in the current stage. The degradation detection and reset module is used to compare the mean values ​​of the first and last segments of the proxy signal sliding window. When the most recent part deteriorates significantly, it determines that the slot has experienced tenant remapping, data distribution drift, or other degradation events, and performs historical statistics clearing and status reset.

[0015] As a preferred option, it also includes: The shared base model module is used to deploy a shared large language model base in edge nodes to host parameter-efficient adapters for multiple tenants; The GPU active adapter slot module is used to maintain K active slots in the GPU memory. Each slot can host at most one tenant adapter and its training state at any given time. The adapter state storage module is used to store inactive tenant adapters, optimizer states, and historical intermediate states in CPU memory and / or SSD, for recovery or swapping out during slot switching.

[0016] Preferably, it also includes: a slot refresh interface module, which can send a refresh suggestion to the system layer when saturation detection is triggered, and the system layer decides whether to reload the new tenant adapter in the slot or reset the status of the slot.

[0017] As a preferred option, for each slot Maintenance latest The sliding window consists of incremental returns. When the window is full, calculate its mean: If the following conditions are met: Then it is determined that the slot has converged or saturated in the current segment.

[0018] As a preferred option, for each slot Maintenance latest Sliding window composed of secondary agent signals When the window is full, use the previous method. One sample was used as a reference interval, and then... Let 1 sample be the current interval; let the sub-window formed by the previous samples be denoted as . The sub-window composed of the latter samples is And calculate the mean of each: When the following conditions are met: Then it is determined that a degradation event has occurred in that slot.

[0019] As a preferred approach, a unified slot reset operation should be defined. ,in For slot numbering, Pick or This is used to indicate the reason for the reset, as follows: like Then set the slot status flag. This indicates that the slot is saturated and will not participate in scheduling again until refreshed at the system level; if Then set This indicates that although the statistics for this slot have been cleared, it remains in a schedulable state. Saturation reset can also send refresh suggestions to the system layer via the interface, requesting the system layer to replace the slot with a new tenant adapter or reset its training state.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: By using GPU-resident adapter slots in edge devices as resource abstractions, and under constraints that GPU memory can only accommodate a limited number of trainable adapters simultaneously and computing power can only execute fine-tuning steps sequentially, this invention performs online computation allocation, convergence identification, degradation detection, and state reset for personalized adapters of multiple tenants. This improves the quality of personalized services for edge-side models and reduces the overhead of ineffective fine-tuning and context switching. This invention can be applied to edge-side intelligent assistants, industry question answering, on-site maintenance, in-vehicle intelligent cockpits, mobile terminal private agents, and large-scale personalized service scenarios requiring localization, low latency, and privacy protection. Attached Figure Description To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 Overall architecture diagram of the adaptive fine-tuning scheduling system for a multi-tenant edge large model to converge perception; Figure 2 This is a schematic diagram of the segmented convergence reward structure; Figure 3 This is a diagram illustrating the state transition of ResetArm. Figure 4 This is a flowchart of an adaptive fine-tuning scheduling method based on slot abstraction. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 like Figure 1 As shown, this invention provides a convergence-aware adaptive fine-tuning scheduling system for large multi-tenant edge models, employing a two-layer decoupled architecture of a "slow-timescale system layer + fast-timescale learning layer". The system layer is responsible for low-frequency tenant adapter admission, replacement, and slot refilling; the learning layer, given the current slot occupancy relationship by the system layer, determines which slot's adapter should be trained next based on the round. To avoid the learning layer directly facing an unbounded set of tenants, this invention abstracts the physical residency locations in the GPU into a fixed number of "adapter slots," treating each slot as a schedulable object. Thus, the learning layer always selects from only a finite number of slots.

[0025] like Figure 2As shown, unlike general non-stationary scheduling, this invention further utilizes the unique "piecewise-converging" structure of multi-tenant edge fine-tuning: within the stable residency phase of the same tenant, continuous fine-tuning typically brings marginal benefits that are initially high and then gradually diminish; while when a slot is occupied by a new tenant or the data distribution of that tenant changes significantly, the slot will enter a new trainable phase. Based on this characteristic, this invention introduces a windowed upper confidence bound decision mechanism and performs two types of lightweight detections simultaneously after each training session: one is convergence / saturation detection based on incremental benefits, and the other is degradation detection based on surrogate signals. The former is used to determine whether a slot is no longer worth training, and the latter is used to determine whether a slot should immediately discard its old history and re-participate in exploration due to environmental changes.

[0026] The multi-tenant edge large model adaptive fine-tuning scheduling system of the present invention includes: Shared Base Model Module: Deploy a shared large language model base in the edge node to host parameter-efficient adapters for multiple tenants.

[0027] GPU Active Adapter Slot Module: Maintains K active slots in GPU memory, with each slot hosting at most one tenant adapter and its training state at any given time.

[0028] Adapter State Storage Module: Stores inactive tenant adapters, optimizer states, and historical intermediate states in CPU memory and / or SSD for recovery or swapping out during slot switching.

[0029] System-level admission / exit module: Based on factors such as cache pressure, service priority, tenant arrival / departure, etc., it determines which tenant adapters enter or leave the GPU slot on a slow time scale and forms a mapping relationship between the current slot and the tenant.

[0030] Learning layer scheduling module: In each fast round, based on the most recent reward statistics and confidence upper bound of each slot, select a slot to perform an incremental fine-tuning.

[0031] Benefit Observation Module: After a fine-tuning step is completed in the selected slot, the incremental performance benefit brought by that step is extracted. For example, the amount of training loss reduction, or equivalent quantities related to the improvement of task quality.

[0032] Agent monitoring module: Extracts lightweight agent signals The preferred method is the instantaneous mini-batch loss that naturally occurs during training, but it can also be the prediction entropy or other statistics that can be obtained without additional verification overhead.

[0033] Convergence perception detection module: Performs a moving average of the most recent incremental gains. When the average value falls below a set threshold, it determines that the slot has converged or saturated in the current stage.

[0034] Degradation Detection and Reset Module: Compares the mean values ​​of the first and last segments of the agent signal sliding window. When the most recent segment deteriorates significantly, it determines that the slot has experienced tenant remapping, data distribution drift, or other degradation events, and performs historical statistics clearing and status reset.

[0035] Slot refresh interface module: When saturation detection is triggered, a refresh suggestion can be sent to the system layer, which will then decide whether to reload the new tenant adapter for the slot or reset the slot's status.

[0036] Furthermore, let the set of physical adapter slots currently available for training at the edge node be: in, This represents the maximum number of tenant adapters that can simultaneously remain in a "directly trainable" state for GPU memory. Let the tenant set be denoted as . Its scale can be much larger than And it can change over time.

[0037] At global round t, the system layer provides an exogenous mapping from slots to tenants: in, Indicates time Occupy slot The tenant. This mapping is determined by the admission / exit policy of the system layer, which is not directly controlled by the learning layer, but is only regarded as the external environment.

[0038] In each round, the learning layer selects a slot. Belongs to slot set Perform an incremental fine-tuning step on the tenant adapter in that slot. After the fine-tuning is complete, the system gains incremental performance benefits. and proxy signals To reflect the fixed cost of each fine-tuning performed on the edge side, this invention introduces an adaptive cost parameter. And define the net reward as: The system optimization objective is to maximize the cumulative net reward within a given time frame T, i.e.: To characterize the evolution of adapter training gains within a slot, this invention divides the time axis of each slot i into several segments: No. The corresponding segment is from arrive The previous time interval is denoted as the th. Each segment is a stable segment. Within the same segment, slots typically correspond to the same tenant or the same stable environment; while different segments correspond to tenant replacement, data migration, or other environmental changes. For slots... In the section The Middle The expected reward for the first time during training is denoted as: This invention assumes that the expected value increases with the number of local training iterations rather than increasing, reflecting the characteristic of diminishing marginal returns.

[0039] For ease of explanation, the piecewise convergence model can be expressed as: at the unknown saturation time... Previously, the expected net reward remained positive and satisfied the lower bound; beyond this point, the expected net reward dropped to a significantly low region, thus becoming reliably detectable. Correspondingly, the following relationship is preferably satisfied: Meanwhile, the proxy signal has a relatively stable mean value within each stable segment. When the environment degrades, its mean increases by at least one detectable interval, i.e., preferably satisfying the following: here, , , These are positive parameters related to the slot or system settings. The above expressions are not required to be precisely known, but rather illustrate how this invention utilizes this structural principle to design an online detection and scheduling mechanism.

[0040] Furthermore, to adapt to non-stationary environments, this invention does not use the historical average reward, but estimates the current slot value only based on the most recent reward sample. Let the reward window length be W, then the set of most recent reward indices for slot i in round t is denoted as: The corresponding window experience average is: in, This represents the number of valid window samples. Further, we define the exploration reward term in fixed confidence form: in, To explore coefficients, This is a confidence parameter. Therefore, the slot... The scheduling index is defined as: If the slot has not yet been sampled, i.e. If the exponent is positive infinity, it is used to ensure the necessary exploration during the cold start phase. Preferably, before formally adopting the above exponent, one or more rounds of polling initialization can be performed on each slot to obtain the initial samples.

[0041] Furthermore, for each slot This invention maintains the latest The sliding window consists of incremental returns. When the window is full, calculate its mean: If the following conditions are met: If the slot has converged or saturated within the current segment, then the marginal benefit of continuing training on it is insufficient to cover the cost of single-step fine-tuning. Therefore, this invention immediately performs a saturation reset operation on the slot and marks it as "temporarily not participating in scheduling" until the system layer refreshes, swaps it out, or reloads it. Through this mechanism, this invention can explicitly distinguish between two situations: "low returns because the prospect of convergence is poor" and "low returns because the learning is almost complete."

[0042] Furthermore, for each slot This invention maintains the latest Sliding window composed of secondary agent signals When the window is full, split it into a preceding reference interval and a following current interval, preferably using the preceding reference interval. One sample was used as a reference interval, and then... Let be the current interval, and let the sub-window formed by the previous samples be denoted as . The sub-window composed of the latter samples is And calculate the mean of each: When the following conditions are met: If a degradation event is detected, the slot is determined to have experienced a degradation event. This degradation event can be: the same slot being replaced by a new tenant adapter at the system layer; a sudden change in the task distribution of the same tenant; a change in training sample attributes due to user switching; or other situations that render old statistical information unreliable. Upon detecting degradation, this invention immediately discards the old rewards and old agent history of the slot and desaturates it, allowing it to re-participate in exploration and adaptive training.

[0043] Furthermore, such as Figure 3 As shown, this invention defines a unified slot reset operation. ,in For slot numbering, Pick or This is used to indicate the reason for the reset. A preferred implementation is as follows: like Then set the slot status flag. This indicates that the slot is saturated and will not participate in scheduling again until refreshed at the system level; if Then set This indicates that although the statistics for the slot have been cleared, it remains schedulable for relearning in the new environment. Preferably, the saturation reset can also send a refresh suggestion to the system layer via an interface, requesting the system layer to replace the slot with a new tenant adapter or reset its training state.

[0044] The present invention has the following advantages and beneficial effects: 1) Decouple the “system layer tenant admission / exit” and “learning layer high-frequency training scheduling” in the continuous fine-tuning process of multi-tenant edge, and establish a limited action space through a fixed number of GPU slots, which significantly reduces the complexity of online decision-making and facilitates integration with existing edge adapter management frameworks.

[0045] 2) Introducing incremental performance benefits explicitly Adaptation Costs with net reward The design allows the scheduling objective to simultaneously reflect model quality improvement and marginal training costs, rather than just pursuing short-term loss reduction.

[0046] 3) Through the convergence / saturation detection mechanism, training can be stopped in time when the tenant adapter has basically learned the skills, thus avoiding wasting limited edge computing power on targets with extremely low marginal returns.

[0047] 4) Through the proxy signal degradation detection and reset mechanism, old history can be quickly discarded after tenant replacement, user handover or data distribution drift occurs, and the ability to adapt to the new environment can be restored, thereby overcoming the problem of slow response of traditional passive forgetting methods.

[0048] 5) By temporarily freezing saturated slots and issuing refresh suggestions to the system layer when necessary, it is possible to reduce the repeated oscillation between multiple converged or low-value slots, significantly reduce context switching and state swapping overhead, and improve system stability.

[0049] 6) The method of this invention does not rely on the system layer to explicitly send complex notifications. The learning layer only needs to read the current slot occupancy, training gains, and agent signals to operate, thus exhibiting good pluggability and engineering deployability.

[0050] 7) Theoretical analysis of this method shows that, under certain conditions, its dynamic regret can grow logarithmically; actual tests also show that this invention outperforms many strong baseline schemes in terms of both model perplexity and switching overhead.

[0051] Example: In this embodiment, the edge node is configured with an NVIDIA RTX 4090 GPU and deploys the shared pedestal model Qwen2.5-1.5B-Instruct. Personalized fine-tuning uses standard LoRA with a precision of BF16, and the LoRA parameters are preferably set to r=8 and alpha=16. The system maintains K=5 active adapter slots in the GPU, with each slot hosting at most one tenant adapter at any given time; the remaining inactive adapters and optimizer states are stored in CPU memory and / or SSD. The system layer uses an LRU-like admission / exit mechanism, updating the current slot mapping based on tenant activity and slot pressure.

[0052] In this embodiment, several tenant data distributions are constructed from UltraChat or other multi-domain dialogue datasets, such as code generation, mathematical reasoning, creative writing, and general question answering. Each physical slot serves a tenant for one session duration, for example, for 800 to 3200 training steps. Then, the system layer can switch it to another tenant to simulate tenant arrival / departure and task switching in a real edge environment. The learning layer selects only one slot to perform incremental fine-tuning in each round. The agent signal preferably directly adopts the training loss of this mini-batch, thus eliminating the need to pause training for validation set inference.

[0053] When a tenant adapter is close to convergence under the current data distribution, the loss reduction per unit fine-tuning step gradually decreases, and the mean of the revenue window may fall below lambda_cost-tau_sat. The system then determines that the slot is saturated and temporarily freezes it, waiting for a system-level refresh before resuming scheduling. If the same slot is occupied by a new tenant, or if the data distribution of the same tenant undergoes a sudden change, the recent agent signal will deteriorate significantly, triggering a degradation reset, causing the slot to re-participate in exploration under the new environment.

[0054] In the tests corresponding to this embodiment, compared with various strong baseline methods, the proposed solution can simultaneously improve model quality and reduce system switching overhead. For example, the geometric mean perplexity (PPL) can be improved by approximately 2.6% to 8.6% relative to the baseline, while the number of training context switches can be reduced by nearly 6 times; compared with the strong non-stationary baseline GLR-klUCB, the number of switches can be reduced from 1609 to 278, a reduction of approximately 82.7%. This demonstrates that the proposed solution can not only more effectively allocate training opportunities to tenants with learning potential, but also significantly suppress unnecessary switching.

[0055] Without increasing additional verification and inference overhead, proxy signals In addition to the preferred use of instantaneous training loss, predictive entropy, normalized gradient statistics, or other lightweight metrics that change in line with the current training state can also be selected. Any metric that exhibits a statistically detectable deterioration trend when a degradation event occurs can be included within the scope of this invention.

[0056] In some applications, the system can additionally set up a "no-op arm," indicating that the adapter update is skipped in the current round, and services are provided only using the existing adapter and shared base. The net reward for this no-op arm can be defined as 0, used to proactively abandon the current training round when all slots have converged or are not worth updating. This extension does not change the core idea of ​​the invention.

[0057] Example 2 like Figure 4 As shown, the present invention also provides a convergence-aware adaptive fine-tuning scheduling method for multi-tenant edge large models, comprising: Step S1: System initialization. Set the number of slots. Rewards Window Convergence detection window Degradation detection window Exploration coefficient Confidence parameters Single-step adaptation cost Convergence threshold and degradation threshold Initialize the counter for each slot. Status flags Historical tenant identifier And clear the reward window, income window and agent window.

[0058] Step S2: In round Initially, the current slot occupancy mapping is read from the system layer. If the current tenant of a slot is found to be the same as the previous record... Inconsistency indicates that tenant remapping or slot refilling has occurred at the system level; in this case, updating is necessary. It will then immediately perform a degradation reset on the slot to clear the old history.

[0059] Step S3: Calculate the scheduling index for all slots. For slots that have been marked as saturated, set their index to negative infinity; for slots without samples, set their index to positive infinity; for normal slots, calculate the index as the sum of the window mean and the exploration term.

[0060] Step S4: Select the slot with the highest index. For the tenants Perform one incremental fine-tuning step.

[0061] Step S5: Extract incremental gains from this training step With agent signal ,according to Calculate net rewards and update the recent rewards window, earnings window, and agent window.

[0062] Step S6: If the profit window is full, check if the convergence / saturation conditions are met; if they are met, immediately perform a saturation reset on the selected slot and temporarily freeze it.

[0063] Step S7: If the agent window is full, compare the average of its front and back segments to determine if a degradation event has occurred; if the degradation conditions are met, immediately perform a degradation reset on the selected slot.

[0064] Step S8: Proceed to the next round, repeating the above process until the set number of rounds is reached or the system stops running.

[0065] The pseudocode expression of the method of this invention is as follows: Algorithm 1: WCA-UCB Main Loop enter: .

[0066] Initialization: for all slots , place , , , .

[0067] for Repeat the following steps: (a) Observe the current tenant mapping for all slots. .

[0068] (b) If Then update and execute .

[0069] (c) Calculate the scheduling index for all slots .

[0070] (d) Choice .

[0071] (e) in the slot Perform one incremental fine-tuning step.

[0072] (f) Observation and and calculate .

[0073] (g) Update , , .

[0074] (h) If Then execute .

[0075] (i) If Then execute .

[0076] Algorithm 2: Will and will Clear as .

[0077] like Then place .

[0078] Otherwise set .

[0079] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A convergent-aware adaptive fine-tuning scheduling method for large multi-tenant edge models, characterized in that, include: Step S1: System initialization, setting the number of slots Rewards Window Convergence detection window Degradation detection window Exploration coefficient Confidence parameters Single-step adaptation cost Convergence threshold and degradation threshold ; Initialize the counter for each slot. Status flags Historical tenant identifier And clear the reward window, earnings window and agent window; Step S2: In round Initially, the current slot occupancy mapping is read from the system layer. ; If the current tenant of a slot is found to be the same as the previous record... Inconsistency indicates that tenant remapping or slot refilling has occurred at the system level; in this case, an update is needed. And immediately perform a degradation reset on the slot to clear the old history; Step S3: Calculate the scheduling index for all slots; for slots that have been marked as saturated, set their index to negative infinity; for slots that have no samples yet, set their index to positive infinity; for normal slots, calculate the index by summing the window mean and the exploration items. Step S4: Select the slot with the highest index. For the tenants Perform one incremental fine-tuning step; Step S5: Extract incremental gains from this training step With agent signal ,according to Calculate net rewards and update the recent rewards window, earnings window, and agent window; Step S6: If the profit window is full, check if the convergence / saturation conditions are met; if they are met, immediately perform a saturation reset on the selected slot and temporarily freeze it. Step S7: If the agent window is full, compare the average of its front and back segments to determine if a degradation event has occurred; if the degradation conditions are met, immediately perform a degradation reset on the selected slot. Step S8: Proceed to the next round, repeating the above process until the set number of rounds is reached or the system stops running.

2. A convergent-aware, multi-tenant edge large-scale adaptive fine-tuning scheduling system, characterized in that, include: The system-level admission / exit module is used to determine whether a tenant adapter enters or leaves a GPU slot based on cache pressure, service priority, and tenant arrival / departure factors on a slow time scale, and to form a mapping relationship between the current slot and the tenant. The learning layer scheduling module is used to select a slot for incremental fine-tuning in each fast round, based on the most recent reward statistics and confidence upper bound of each slot. The benefit observation module is used to extract the incremental performance benefit brought by the fine-tuning step after a fine-tuning step is completed in the selected slot. ; The agent monitoring module is used to extract lightweight agent signals. ; The convergence perception detection module is used to perform a moving average of the recent incremental gains. When the moving average is lower than a set threshold, it is determined that the slot has converged or saturated in the current stage. The degradation detection and reset module is used to compare the mean values ​​of the first and last segments of the proxy signal sliding window. When the most recent part deteriorates significantly, it determines that the slot has experienced tenant remapping, data distribution drift, or other degradation events, and performs historical statistics clearing and status reset.

3. The convergence-aware multi-tenant edge large model adaptive fine-tuning scheduling system as described in claim 2, characterized in that, Also includes: The shared base model module is used to deploy a shared large language model base in edge nodes to host parameter-efficient adapters for multiple tenants; The GPU active adapter slot module is used to maintain K active slots in the GPU memory. Each slot can host at most one tenant adapter and its training state at any given time. The adapter state storage module is used to store inactive tenant adapters, optimizer states, and historical intermediate states in CPU memory and / or SSD, for recovery or swapping out during slot switching.

4. The convergence-aware multi-tenant edge large model adaptive fine-tuning scheduling system as described in claim 3, characterized in that, Also includes: The slot refresh interface module is used to send a refresh suggestion to the system layer when saturation detection is triggered. The system layer then decides whether to reload the new tenant adapter for the slot or reset the slot's state.

5. The convergence-aware multi-tenant edge large model adaptive fine-tuning scheduling system as described in claim 4, characterized in that, For each slot Maintenance latest The sliding window consists of incremental returns. When the window is full, calculate its mean: If the following conditions are met: Then it is determined that the slot has converged or saturated in the current segment.

6. The convergence-aware multi-tenant edge large model adaptive fine-tuning scheduling system as described in claim 5, characterized in that, For each slot Maintenance latest Sliding window composed of secondary agent signals ; When the window is full, use the previous method. One sample was used as a reference interval, and then... Let 1 sample be the current interval; let the sub-window formed by the previous samples be denoted as . The sub-window composed of the latter samples is And calculate the mean of each: When the following conditions are met: If so, it is determined that a degradation event has occurred in that slot.

7. The convergence-aware multi-tenant edge large model adaptive fine-tuning scheduling system as described in claim 6, characterized in that, Define a unified slot reset operation ,in For slot numbering, Pick or This is used to indicate the reason for the reset, as follows: like Then set the slot status flag. This indicates that the slot is saturated and will not participate in scheduling again until refreshed at the system level; if Then set This indicates that although the statistics for this slot have been cleared, it remains in a schedulable state. Saturation reset can also send refresh suggestions to the system layer via the interface, requesting the system layer to replace the slot with a new tenant adapter or reset its training state.

Images