Large model adaptive fusion method and system based on deep learning

Abstract

The invention discloses a large model adaptive fusion method and system based on deep learning, and the method comprises the following steps: obtaining input data and a task identifier, and configuring a large model set, an adaptation layer set and a low-rank increment parameter structure; encoding input data to obtain a context vector, and generating a model-level rank budget, a hierarchical rank upper limit and a hierarchical rank allocation table; executing improved AdaLoRA training according to a hierarchical rank distribution table and updating an effective rank; calculating a target index by using verification data, updating a model-level rank budget and a hierarchical rank upper limit, and updating a hierarchical rank distribution table; converting the low-rank increment parameter structure into subspace representation and generating a conflict mark; performing orthogonal complementary projection and rank clipping on the subspace representation corresponding to the conflict mark to obtain a release rank budget, redistributing the release rank budget, and updating a hierarchical rank distribution table; outputting a final hierarchical rank distribution table and an effective rank; and the reasoning stage is conditionally activated, fused and output. The stability and consistency of fusion output are improved under the constraint of the total rank budget.

Description

Technical Field

[0001] This invention relates to the field of deep learning and artificial intelligence model fusion technology, and in particular to a large-scale adaptive fusion method and system based on deep learning. Background Technology

[0002] In recent years, large models have been widely used in tasks such as natural language processing, computer vision, speech recognition, and code generation. The industry often uses pre-trained base models to fine-tune them with task data to adapt to different business scenarios. To reduce fine-tuning costs and storage overhead, efficient parameter fine-tuning techniques have gradually become the mainstream approach. A typical practice is to introduce low-rank incremental parameter structures into the attention projection linear layer and feedforward network linear layer of the large model, training only the incremental parameters while keeping the original model parameters stable, thereby achieving multi-task deployment under limited computing power. At the same time, for complex tasks and multi-domain data, common systems improve coverage by building large model sets and introduce fusion strategies in the inference stage to alleviate single-model bias. Fusion methods include weighted fusion, gating selection, and dynamic routing, in order to obtain more stable outputs under different input distributions.

[0003] Existing efficient parameter fine-tuning and model fusion solutions face challenges in resource allocation and collaborative control in large model ensemble scenarios. On the one hand, low-rank fine-tuning often relies on fixed-rank configuration or independent rank scheduling at a single layer, making it difficult to allocate rank budgets synchronously at the model and layer levels based on input data and task identifiers. This results in inconsistent resource allocation under the total rank budget constraint, leading to situations where some models or adaptation layers obtain too much effective rank while other positions lack sufficient effective rank. On the other hand, dynamic rank adjustment largely depends on local importance assessment driven by training loss, lacking a feedback mechanism for target indicators on validation data. This makes it difficult to perform meta-updates on the model-level rank budget and the upper limit of layer rank, resulting in the layer rank allocation failing to converge stably with changes in task difficulty and data distribution. The rank evolution during training is prone to frequent fluctuations, affecting training controllability and cross-model comparability.

[0004] In scenarios involving parallel multi-adapter layers and multi-model fusion, existing technologies still suffer from difficulties in measuring and constraining interference between adapter layers. Low-rank incremental parameters of different adapter layers may learn highly similar subspace directions in the representation space, leading to information redundancy and parameter conflicts. This manifests as inconsistent offsets in the outputs of different models on similar inputs, thereby weakening the stability of fusion. Common fusion strategies mainly rely on output layer weighting, lacking conflict labeling and orthogonal constraint mechanisms for incremental subspaces. They cannot perform orthogonal projection and rank pruning on conflict subspaces during the training phase, nor can they redistribute the released rank budget according to marginal returns. Thus, it is difficult to simultaneously consider differentiated representations and fusion controllability under the constraint of total rank budget. Furthermore, the calculation of fusion coefficients during the inference phase lacks a unified rule that links with conflict information.

[0005] Therefore, how to provide a large-scale adaptive fusion method and system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based adaptive fusion method and system for large models. This invention introduces a low-rank incremental parameter structure into a large model set and combines it with an improved AdaLoRA training mechanism, hierarchical rank allocation, and an adaptive fusion strategy based on conflict constraints to achieve collaborative training and inference fusion of different models under a unified rank budget constraint. By utilizing context-driven rank allocation, validation metric-guided meta-updates, and subspace conflict suppression and rank reassignment mechanisms, effective collaboration and adaptive fusion between large models are achieved. This method has the advantages of controllable parameter overhead, stable training process, and strong consistency of fusion results.

[0007] According to an embodiment of the present invention, a large model adaptive fusion method based on deep learning includes the following steps:

[0008] Obtain input data and task identifiers, determine the large model set and the adaptation layer set, and establish a low-rank incremental parameter structure for each adaptation layer in the adaptation layer set;

[0009] The input data is encoded to obtain a context vector. The model-level rank budget and the hierarchical rank upper limit are generated based on the context vector. The hierarchical rank allocation table is generated based on the model-level rank budget, the hierarchical rank upper limit, and the total rank budget.

[0010] The improved AdaLoRA training and update is performed based on the hierarchical rank allocation table, which trains the low-rank incremental parameter structure and dynamically adjusts the effective rank.

[0011] Calculate the target indicators based on the validation data, perform meta-updates on the model-level rank budget and the hierarchical rank upper limit based on the target indicators, and update the hierarchical rank allocation table based on the meta-update results.

[0012] The low-rank incremental parameter structure is converted into a subspace representation, the distance between subspaces is calculated, and a conflict marker is generated based on the distance between subspaces.

[0013] Perform orthogonal complementary projection on the subspace representation corresponding to the conflict marker, perform rank pruning to obtain the release rank budget, sort and redistribute the release rank budget based on marginal revenue, and update the hierarchical rank allocation table;

[0014] Output the final hierarchical rank allocation table and effective rank based on the preset termination conditions;

[0015] During the inference phase, conditional activation is performed based on the final hierarchical rank allocation table, fusion coefficients are generated based on marginal returns and conflict markers, and adaptive fusion is performed on the output of the large model set based on the fusion coefficients, and the fusion result is output.

[0016] Optionally, determining the large model set and the adaptation layer set, and establishing a low-rank incremental parameter structure for each adaptation layer in the adaptation layer set, specifically includes:

[0017] The set of large models and the set of adaptation layers are determined based on the task identifier. The set of large models contains at least two large models, and each large model corresponds to a model identifier.

[0018] For each large model in the large model set, an adaptation layer set is obtained by filtering based on a preset layer type set. The preset layer type set includes attention projection linear layers and feedforward network linear layers.

[0019] For each adaptation layer in the adaptation layer set, a low-rank incremental parameter structure is established. The low-rank incremental parameter structure includes a first low-rank parameter matrix and a second low-rank parameter matrix. The connection dimension between the first low-rank parameter matrix and the second low-rank parameter matrix is ​​used as the effective rank.

[0020] The updated weight matrix is ​​constructed based on the adaptation layer weight matrix, the low-rank incremental parameter structure, and the scaling factor. The scaling factor is set before training begins and corresponds one-to-one with the adaptation layer identifier.

[0021] Optionally, the generation of the hierarchical rank allocation table specifically includes:

[0022] The input data is encoded to obtain a context vector;

[0023] The model-level rank budget and the hierarchical rank upper limit are generated based on the context vector. The model-level rank budget corresponds one-to-one with the large model set dimension, and the hierarchical rank upper limit corresponds one-to-one with the adaptation layer set dimension.

[0024] Under the total rank budget constraint, a hierarchical rank allocation table is constructed. The hierarchical rank allocation table is a two-dimensional index structure composed of the large model identifier and the adaptation layer identifier. Each index position corresponds to a valid rank value.

[0025] Based on the model-level rank budget constraint, the hierarchical rank allocation table is allocated according to the major model dimension, so that the sum of the effective ranks corresponding to each major model is equal to the model-level rank budget.

[0026] Based on the upper limit constraint of hierarchical rank, the hierarchical rank allocation table is pruned according to the adaptation layer dimension so that the effective rank corresponding to each adaptation layer does not exceed the upper limit of hierarchical rank.

[0027] Perform a global correction on the hierarchical rank allocation table after allocation and trimming, so that the sum of all effective ranks in the hierarchical rank allocation table equals the total rank budget, and complete the generation of the hierarchical rank allocation table.

[0028] Optionally, the definition and training update of the improved AdaLoRA specifically include:

[0029] The effective rank of each adaptation layer in the adaptation layer set is determined based on the hierarchical rank allocation table. Structural constraints are applied to the low-rank incremental parameter structure based on the effective rank, so that the number of rank components participating in training in the low-rank incremental parameter structure is consistent with the effective rank.

[0030] Based on the low-rank incremental parameter structure, an improved AdaLoRA training mechanism is used to perform training updates on the adaptation layer set. The improved AdaLoRA training mechanism includes:

[0031] During training, parameter updates are performed only on the rank components corresponding to the valid ranks indicated by the hierarchical rank allocation table, and freezing control is performed on the rank components not indicated by the hierarchical rank allocation table, so that the parameters of the frozen rank components remain unchanged during training.

[0032] Within a preset training period, the rank importance score is calculated for each rank component in the low-rank incremental parameter structure based on the training loss. The rank importance score is determined jointly based on the gradient magnitude and parameter magnitude of the corresponding rank component.

[0033] Rank evolution control is performed based on rank importance scores and hierarchical rank allocation tables. When the rank importance score is lower than the first threshold, a pruning operation is performed on the corresponding rank component. When the rank importance score is higher than the second threshold and there is an available rank budget in the hierarchical rank allocation table, an addition operation is performed on the corresponding rank component.

[0034] During the rank evolution control process, the change magnitude constraints are applied to the pruning and addition operations to ensure that the number of effective rank changes of the adaptation layer within a single training cycle does not exceed the preset change upper limit, and the effective rank in the low-rank incremental parameter structure is updated synchronously after the change is completed.

[0035] Optionally, the step of updating the hierarchical rank allocation table based on the meta-update result specifically includes:

[0036] A verification batch is constructed based on the verification data. Conditional activation is performed on the adaptation layer set based on the hierarchical rank allocation table. Verification inference is performed on the large model set after conditional activation to obtain the verification output.

[0037] The verification loss is calculated based on the verification output and the verification label. The verification loss is the function value of the verification output and the verification label under the preset loss function.

[0038] The target indicator is calculated based on the verification loss. The target indicator is obtained by combining the verification loss and the total rank budget penalty term. The total rank budget penalty term is determined by the difference between the sum of all effective ranks in the hierarchical rank allocation table and the total rank budget. The target indicator corresponds one-to-one with the task identifier.

[0039] Meta-updates are performed on the model-level rank budget and the hierarchical rank upper limit based on the target metric. The meta-update includes calculating the sensitivity of the target metric to the model-level rank budget and the hierarchical rank upper limit and generating an update increment. The updated model-level rank budget and the updated hierarchical rank upper limit are then updated based on the update increment to obtain the updated model-level rank budget and the updated hierarchical rank upper limit.

[0040] The hierarchical rank allocation table is updated based on the updated model-level rank budget, the updated hierarchical rank upper limit, and the total rank budget. This ensures that the sum of all valid ranks in the hierarchical rank allocation table equals the total rank budget, that the valid rank corresponding to each adaptation layer in the hierarchical rank allocation table does not exceed the updated hierarchical rank upper limit, and that the sum of the valid ranks corresponding to each major model in the hierarchical rank allocation table equals the updated model-level rank budget.

[0041] Optionally, the generation of the conflict marker specifically includes:

[0042] Based on the adaptation layer set, the low-rank incremental parameter structure is traversed, and for each adaptation layer in the adaptation layer set, the low-rank incremental parameter structure is converted into a subspace representation.

[0043] The subspace representation is determined by the first low-rank parameter matrix and the second low-rank parameter matrix. The subspace representation contains effective rank basis vectors. The basis vectors are determined by the column vectors of the first low-rank parameter matrix. There is a one-to-one correspondence between the basis vectors and the effective ranks.

[0044] Construct a set of subspace pairs in the dimension of the adapter layer set. The set of subspace pairs contains a pair of adapter layer identifiers and a pair of subspace representations. The adapter layer identifier pair consists of two different adapter layer identifiers in the adapter layer set, and the subspace representation pair consists of two subspace representations corresponding to the adapter layer identifier pair.

[0045] For each pair of subspaces, calculate the distance between subspaces, which is obtained by performing a distance metric operation on the set of basis vectors representing the pair of subspaces;

[0046] Conflict markers are generated based on the distance between subspaces. When the distance between subspaces is less than the distance threshold, conflict markers are generated for the adaptation layer identifier pairs. When the distance between subspaces is greater than the distance threshold, no conflict markers are generated. The distance threshold is set before training begins and corresponds one-to-one with the task identifier.

[0047] Optionally, the step of sorting and redistributing the release rank budget based on marginal revenue and updating the hierarchical rank allocation table specifically includes:

[0048] The set of adaptation layer identifier pairs is determined based on the conflict markers. The set of adaptation layer identifier pairs consists of the adaptation layer identifier pairs corresponding to the conflict markers.

[0049] For each adaptation layer identifier pair in the adaptation layer identifier pair set, read the corresponding subspace representation pair of the adaptation layer identifier pair, perform orthogonal complement projection on each subspace representation in the subspace representation pair, and project the subspace representation onto the orthogonal complement space of another subspace representation in the subspace representation pair, and output the projected subspace representation;

[0050] Rank pruning is performed based on the projected subspace representation. Rank pruning includes determining the set of rank components participating in training in the projected subspace representation, removing rank components that do not enter the set of rank components from the effective rank, and determining the released rank budget based on the difference between the effective rank before rank pruning and the effective rank after rank pruning.

[0051] The release rank budget is sorted and redistributed based on marginal revenue. Marginal revenue is calculated on the validation data. Marginal revenue is the negative of the change in the target indicator when the effective rank corresponding to the adapter layer identifier in the adapter layer set increases by one unit. The change in the target indicator is determined by the difference between the target indicator before increasing the effective rank by one unit and the target indicator after increasing the effective rank by one unit.

[0052] The release rank budget is reallocated based on the marginal revenue ranking results. The reallocation includes allocating effective rank increments to the adaptation layer identifiers in the adaptation layer set in descending order of marginal revenue, and updating the hierarchical rank allocation table based on the allocation results.

[0053] Optionally, the output of the fusion result specifically includes:

[0054] Read the final hierarchical rank allocation table and perform conditional activation on the adaptation layer set;

[0055] The large model set is invoked to perform inference on the input data, and the output of each large model in the large model set is obtained.

[0056] For each large model in the large model set, the marginal benefit value and the conflict mark count value are read. The marginal benefit value and the conflict mark count value are converted into a model score according to the preset calculation rules. The model score is then normalized according to the large model set dimension to obtain the fusion coefficient.

[0057] Adaptive fusion is performed on the model output according to the fusion coefficient, and the fusion result is output.

[0058] According to an embodiment of the present invention, a large-model adaptive fusion system based on deep learning includes:

[0059] The data and task acquisition module is used to acquire input data and task identifiers;

[0060] The configuration module is used to determine the large model set and the adaptation layer set based on the task identifier, and to establish a low-rank incremental parameter structure for each adaptation layer in the adaptation layer set.

[0061] The hierarchical rank allocation table generation module is used to encode the input data to obtain the context vector, generate the model-level rank budget and the hierarchical rank upper limit, and generate the hierarchical rank allocation table based on the model-level rank budget, the hierarchical rank upper limit, and the total rank budget.

[0062] The training update module is used to perform improved AdaLoRA training updates based on the hierarchical rank allocation table, train the low-rank incremental parameter structure and dynamically adjust the effective rank.

[0063] The meta-update module is used to calculate the target index based on the validation data, perform meta-updates on the model-level rank budget and the hierarchical rank upper limit based on the target index, and update the hierarchical rank allocation table based on the meta-update results.

[0064] The conflict marking module is used to convert the low-rank incremental parameter structure into a subspace representation, calculate the distance between subspaces, and generate conflict markers based on the distance between subspaces.

[0065] The rank redistribution module is used to perform orthogonal complementary projection on the subspace representation corresponding to the conflict marker, perform rank pruning to obtain the release rank budget, sort and redistribute the release rank budget based on marginal benefits, and update the hierarchical rank allocation table.

[0066] The fusion inference module is used to perform conditional activation based on the final hierarchical rank allocation table, generate fusion coefficients based on marginal benefit values ​​and conflict marker counts, and perform adaptive fusion on the output of the large model set based on the fusion coefficients to output the fusion result.

[0067] The beneficial effects of this invention are:

[0068] This invention generates model-level rank budget and hierarchical rank upper limit through context vector-driven model-level rank budget and forms a hierarchical rank allocation table under the constraint of total rank budget. This enables large model sets to obtain a consistent resource allocation entry point during the training phase. Combined with the target index construction and meta-update mechanism on validation data, the model-level rank budget and hierarchical rank upper limit are adaptively adjusted with training feedback, and the hierarchical rank allocation table is iteratively corrected. This makes the allocation and adjustment process of effective rank have a controllable convergence path, reduces resource mismatch caused by fixed rank configuration, and improves adaptability under conditions of task changes and input distribution changes.

[0069] This invention improves AdaLoRA training updates by introducing effective rank constraints, freeze control, rank importance scoring, and restricted rank evolution. This ensures that the training updates and dynamic adjustments of low-rank incremental parameter structures are uniformly constrained by the hierarchical rank allocation table, reducing invalid rank updates and training fluctuations. At the subspace level, the low-rank incremental parameter structure is converted into a subspace representation. Conflict markers are generated using the distance between subspaces, and orthogonal complementary projection and rank pruning are performed on the subspace representations corresponding to the conflict markers to release the rank budget. Then, the hierarchical rank allocation table is reassigned and updated according to marginal returns, maintaining the distinction between the representation directions of different adaptation layers and suppressing subspace redundancy and parameter conflicts. During the inference phase, conditional activation is performed based on the final hierarchical rank allocation table, and fusion coefficients are generated by combining marginal returns and conflict markers. This achieves adaptive fusion of the outputs of large model sets, improving the consistency and stability of the fused outputs, while maintaining the parameter size constrained by the total rank budget. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is a flowchart of a large-model adaptive fusion method based on deep learning proposed in this invention;

[0072] Figure 2 This is a schematic diagram illustrating the generation of the hierarchical rank allocation table for a large-model adaptive fusion method based on deep learning proposed in this invention.

[0073] Figure 3 This is a schematic diagram of subspace conflict labeling for a large-model adaptive fusion method based on deep learning proposed in this invention. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0075] refer to Figure 1-3 A deep learning-based adaptive fusion method for large models includes the following steps:

[0076] Obtain input data and task identifiers, determine the large model set and the adaptation layer set, and establish a low-rank incremental parameter structure for each adaptation layer in the adaptation layer set;

[0077] The input data is encoded to obtain a context vector. The model-level rank budget and the hierarchical rank upper limit are generated based on the context vector. The hierarchical rank allocation table is generated based on the model-level rank budget, the hierarchical rank upper limit, and the total rank budget.

[0078] The improved AdaLoRA training and update is performed based on the hierarchical rank allocation table, which trains the low-rank incremental parameter structure and dynamically adjusts the effective rank.

[0079] Calculate the target indicators based on the validation data, perform meta-updates on the model-level rank budget and the hierarchical rank upper limit based on the target indicators, and update the hierarchical rank allocation table based on the meta-update results.

[0080] The low-rank incremental parameter structure is converted into a subspace representation, the distance between subspaces is calculated, and a conflict marker is generated based on the distance between subspaces.

[0081] Perform orthogonal complementary projection on the subspace representation corresponding to the conflict marker, perform rank pruning to obtain the release rank budget, sort and redistribute the release rank budget based on marginal revenue, and update the hierarchical rank allocation table;

[0082] Output the final hierarchical rank allocation table and effective rank based on the preset termination conditions;

[0083] During the inference phase, conditional activation is performed based on the final hierarchical rank allocation table, fusion coefficients are generated based on marginal returns and conflict markers, and adaptive fusion is performed on the output of the large model set based on the fusion coefficients, and the fusion result is output.

[0084] In this embodiment, determining the large model set and the adaptation layer set, and establishing a low-rank incremental parameter structure for each adaptation layer in the adaptation layer set, specifically includes:

[0085] The large model set and the adaptation layer set are determined based on the task identifier. The large model set contains at least two large models, each of which corresponds to a model identifier. The task identifier is used as the search key to match the model identifier set in the model configuration table. The model configuration table provides the correspondence between the model identifier and the model parameter file path, the model network structure description, and the model input and output interface description. Based on the matching results, the model parameter file is loaded and the network structure is instantiated to obtain the large model set.

[0086] For each large model in the large model set, a set of suitable layers is obtained by filtering based on a preset set of layer types. The preset set of layer types includes attention projection linear layers and feedforward linear layers. The filtering process is based on the network structure description to traverse the network layer sequence, read the layer type label and weight matrix attributes of each network layer, and add the network layer to the set of suitable layers if the layer type label belongs to attention projection linear layers or feedforward linear layers. Network layers that do not meet the preset set of layer types are not added to the set of suitable layers.

[0087] For each adaptation layer in the adaptation layer set, a low-rank incremental parameter structure is established. The low-rank incremental parameter structure includes a first low-rank parameter matrix and a second low-rank parameter matrix. The first low-rank parameter matrix and the second low-rank parameter matrix correspond one-to-one with the adaptation layer weight matrix. The row dimension of the first low-rank parameter matrix is ​​consistent with the output dimension of the adaptation layer weight matrix, and the column dimension of the second low-rank parameter matrix is ​​consistent with the input dimension of the adaptation layer weight matrix. The column dimension of the first low-rank parameter matrix and the row dimension of the second low-rank parameter matrix are consistent and are used as the effective rank. The effective rank is used as the upper limit of the trainable degrees of freedom of the low-rank incremental parameter structure. The first low-rank parameter matrix and the second low-rank parameter matrix are initialized as trainable parameters before training begins.

[0088] An updated weight matrix is ​​constructed based on the adaptation layer weight matrix, the low-rank incremental parameter structure, and the scaling factor. The updated weight matrix replaces the adaptation layer weight matrix in the linear transformation calculation during the forward computation. The low-rank incremental parameter structure first calculates the low-rank incremental matrix. The low-rank incremental matrix is ​​multiplied by the scaling factor to obtain the scaling incremental matrix. The scaling incremental matrix is ​​added to the adaptation layer weight matrix to obtain the updated weight matrix. The scaling factor is set before training begins and corresponds one-to-one with the adaptation layer identifier. The adaptation layer identifier is used to index the scaling factor and ensure that the scaling factor remains distinguishable between different adaptation layers.

[0089] In this embodiment, the generation of the hierarchical rank allocation table specifically includes:

[0090] The input data is encoded to obtain a context vector. The encoding process reads the data type identifier of the input data. When the data type identifier is text, a text encoding network is used to output a text feature vector. When the data type identifier is an image, a visual encoding network is used to output a visual feature vector. When the data type identifier is a sequence, a sequence encoding network is used to output a sequence feature vector. The task identifier is input into the task embedding network and outputs a task embedding vector. The context vector is obtained by concatenating the input data feature vector and the task embedding vector according to the vector dimension.

[0091] The model-level rank budget and the hierarchical rank upper limit are generated based on context vectors. The model-level rank budget generation process inputs the context vector into the rank budget to generate the network output model budget weight vector. The model budget weight vector is normalized in the dimension of the large model set to obtain the model budget proportion vector. The model budget proportion vector is multiplied by the total rank budget to obtain the model-level rank budget and is then integerized. The hierarchical rank upper limit generation process inputs the context vector into the hierarchical rank upper limit to generate the network output hierarchical rank upper limit weight vector. The hierarchical rank upper limit weight vector is arranged in the order of the adaptation layer set and is scaled and integerized to obtain the hierarchical rank upper limit.

[0092] Under the total rank budget constraint, a hierarchical rank allocation table is constructed. The hierarchical rank allocation table adopts a two-dimensional mapping structure. The first dimension of the two-dimensional mapping structure is the model identifier, and the second dimension is the adaptation layer identifier. Each position of the two-dimensional mapping structure stores a valid rank value, and the valid rank value is a non-negative integer.

[0093] Based on the model-level rank budget constraint, the hierarchical rank allocation table is allocated according to the large model dimension. During the allocation process, the model-level rank budget is extracted for each model identifier as the model-in-model allocable rank budget. An in-model allocation weight sequence is generated on the adaptation layer identifier sequence corresponding to the model identifier. The in-model allocation weight sequence is composed of the components of the hierarchical upper limit weight vector on the adaptation layer identifier sequence and is normalized. The model-in-model allocable rank budget is discretely allocated according to the in-model allocation weight sequence to obtain a set of effective rank values ​​and written into the hierarchical rank allocation table, so that the sum of the effective rank values ​​of the corresponding row of the model identifier is equal to the model-level rank budget.

[0094] Based on the upper limit constraint of the hierarchical rank, the hierarchical rank allocation table is pruned according to the adaptation layer dimension. The pruning process traverses all positions of the hierarchical rank allocation table. When the effective rank value of any position is greater than the upper limit value of the hierarchical rank at the corresponding adaptation layer identifier, the effective rank value is replaced with the upper limit value, and the pruning difference is accumulated to obtain the rank budget to be recovered.

[0095] A global correction is performed on the hierarchical rank allocation table after allocation and trimming. The global correction process calculates the difference between the sum of all valid rank values ​​in the hierarchical rank allocation table and the total rank budget as the correction difference. When the correction difference is positive, incremental allocation is performed in the set of two-dimensional mapping positions that have not reached the upper limit of hierarchical rank according to the weight sequence allocated in the model until the correction difference is 0. When the correction difference is negative, the reduction and recycling are performed in the set of two-dimensional mapping positions with valid rank values ​​greater than 0 according to the reverse order of the weight sequence allocated in the model until the correction difference is 0, thus completing the generation of the hierarchical rank allocation table.

[0096] In this embodiment, the improved definition and training update of AdaLoRA specifically include:

[0097] The effective rank corresponding to each adaptation layer in the adaptation layer set is determined based on the hierarchical rank allocation table. The determination includes reading the effective rank value corresponding to the adaptation layer identifier in the hierarchical rank allocation table, and applying structural constraints to the low-rank incremental parameter structure based on the effective rank. The structural constraints include limiting the number of elements in the rank component set participating in training in the low-rank incremental parameter structure to be equal to the effective rank, so that the number of rank components participating in training in the low-rank incremental parameter structure is consistent with the effective rank.

[0098] Based on a low-rank incremental parameter structure, an improved AdaLoRA training mechanism is used to perform training updates on the adaptation layer set. The training update includes: performing forward computation based on the input data and the updated weight matrix to obtain the training output; calculating the training loss based on the training output and the supervision signal; and calculating the parameter gradients of the low-rank incremental parameter structure based on the training loss. The improved AdaLoRA training mechanism includes:

[0099] During training, parameter updates are performed only on the rank components corresponding to the valid ranks indicated by the hierarchical rank allocation table. The parameter updates include optimizing the rank components participating in training based on the parameter gradient, and freezing control is performed on the rank components not indicated by the hierarchical rank allocation table. The freezing control includes prohibiting frozen rank components from participating in the optimizer update, so that the parameters of the frozen rank components remain unchanged during training.

[0100] Within a preset training period, a rank importance score is calculated for each rank component in the low-rank incremental parameter structure based on the training loss. The calculation includes generating a rank importance score based on the parameter gradient magnitude and parameter magnitude within the training period. The rank importance score is determined jointly based on the gradient magnitude and parameter magnitude of the corresponding rank component.

[0101] Rank evolution control is performed based on rank importance scores and hierarchical rank allocation tables. When the rank importance score is lower than a first threshold, a pruning operation is performed on the corresponding rank component. The pruning operation includes removing the corresponding rank component from the set of rank components participating in training. When the rank importance score is higher than a second threshold and there is an available rank budget in the hierarchical rank allocation table, an addition operation is performed on the corresponding rank component. The addition operation includes adding the corresponding rank component to the set of rank components participating in training.

[0102] During the rank evolution control process, the pruning and addition operations are subject to change range constraints. The change range constraints include limiting the change in the number of elements in the rank component set participating in training within a preset training period to not exceed a preset change upper limit, so that the number of effective rank changes in the adaptation layer within a single training period does not exceed the preset change upper limit, and the effective rank in the low-rank incremental parameter structure is updated synchronously after the change is completed.

[0103] In this embodiment, updating the hierarchical rank allocation table based on the meta-update result specifically includes:

[0104] A verification batch is constructed based on the verification data. The verification batch consists of verification input and verification label. The hierarchical rank allocation table serves as the control input for verification inference. The verification inference process calls the large model one by one in the large model set. For each large model, the adaptation layer is called one by one in the adaptation layer set. At the adaptation layer, the corresponding low-rank incremental parameter structure is executed according to the effective rank value provided by the hierarchical rank allocation table to participate in the calculation and output the verification output.

[0105] The validation loss is calculated based on the validation output and validation label. The validation loss is obtained by operating on the validation output and validation label using a preset loss function, which is set before training begins.

[0106] The target indicator is calculated based on the verification loss. The target indicator is obtained by combining the verification loss and the total rank budget penalty term. The total rank budget penalty term is determined by the difference between the sum of all effective ranks in the hierarchical rank allocation table and the total rank budget. The target indicator corresponds one-to-one with the task identifier.

[0107] Meta-updates are performed on the model-level rank budget and the hierarchical rank upper limit based on the target metric. The meta-updates include calculating the sensitivity of the target metric to the model-level rank budget and the sensitivity of the target metric to the hierarchical rank upper limit. The sensitivity calculation uses a perturbation method to generate the target metric difference and forms the update increment accordingly. The model-level rank budget and the hierarchical rank upper limit are updated based on the update increment, and the updated model-level rank budget and the updated hierarchical rank upper limit are output.

[0108] The hierarchical rank allocation table is updated based on the updated model-level rank budget, the updated hierarchical rank upper limit, and the total rank budget. The update process ensures that the sum of all valid ranks in the hierarchical rank allocation table equals the total rank budget, that the valid rank corresponding to each adaptation layer in the hierarchical rank allocation table does not exceed the updated hierarchical rank upper limit, and that the sum of the valid ranks corresponding to each major model in the hierarchical rank allocation table equals the updated model-level rank budget.

[0109] In this embodiment, the generation of the conflict marker specifically includes:

[0110] Based on the adaptation layer set, the low-rank incremental parameter structure is traversed. During the traversal process, the identifier of each adaptation layer in the adaptation layer set is read. Based on the adaptation layer identifier, the corresponding low-rank incremental parameter structure is located. The low-rank incremental parameter structure contains a first low-rank parameter matrix and a second low-rank parameter matrix. The first low-rank parameter matrix and the second low-rank parameter matrix are used as inputs for the subspace representation.

[0111] The subspace representation is determined by the first low-rank parameter matrix and the second low-rank parameter matrix. The subspace representation contains effective rank basis vectors. The basis vectors are determined by the column vectors of the first low-rank parameter matrix. The set of basis vectors consists of column vectors with effective rank. Each basis vector in the set of basis vectors corresponds to a rank component. The number of rank components is equal to the effective rank.

[0112] In the adaptation layer set dimension, a subspace pair set is constructed. The subspace pair set construction process selects two different adaptation layer identifiers in the adaptation layer set to form an adaptation layer identifier pair. The two subspace representations corresponding to the adaptation layer identifier pair form a subspace representation pair. The subspace pair set is composed of the combination of adaptation layer identifier pairs and subspace representation pairs.

[0113] For each subspace pair, calculate the distance between subspaces. The process of calculating the distance between subspaces involves performing a distance metric operation on the two basis vector sets in the subspace representation pair. The distance metric operation outputs the distance value between subspaces.

[0114] Conflict markers are generated based on the distance between subspaces. The conflict marker generation process reads the distance threshold, which is set before training begins and corresponds one-to-one with the task identifier. When the distance between subspaces is less than the distance threshold, conflict markers are generated for the adaptation layer identifier pair. When the distance between subspaces is greater than the distance threshold, no conflict markers are generated.

[0115] In this embodiment, the step of sorting and redistributing the release rank budget based on marginal revenue and updating the hierarchical rank allocation table specifically includes:

[0116] The set of adaptation layer identifier pairs is determined based on the conflict marker. There is a one-to-one correspondence between the conflict marker and the adaptation layer identifier pair. The adaptation layer identifier pair consists of two different adaptation layer identifiers in the adaptation layer set. When the conflict marker value is conflict, the corresponding adaptation layer identifier pair is added to the set of adaptation layer identifier pairs.

[0117] For each adaptation layer identifier pair in the adaptation layer identifier pair set, read the corresponding subspace representation pair. The subspace representation pair includes a first subspace representation and a second subspace representation. The first subspace representation corresponds to the first adaptation layer identifier in the adaptation layer identifier pair, and the second subspace representation corresponds to the second adaptation layer identifier in the adaptation layer identifier pair. Perform orthogonal complement projection on the first subspace representation to obtain the first projected subspace representation. The orthogonal complement projection process calculates the orthogonal complement space based on the subspace determined by the second subspace representation and projects the first subspace representation onto the orthogonal complement space. Perform orthogonal complement projection on the second subspace representation to obtain the second projected subspace representation. The orthogonal complement projection process calculates the orthogonal complement space based on the subspace determined by the first subspace representation and projects the second subspace representation onto the orthogonal complement space.

[0118] Rank pruning is performed based on the projected subspace representation. The rank pruning process is performed separately for the first and second projected subspace representations. The rank pruning process calculates the magnitude of the basis vectors in the basis vector set corresponding to the projected subspace representation and determines the set of rank components to participate in training based on the magnitude. The size of the rank component set does not exceed the effective rank before rank pruning. Rank components that do not enter the rank component set are removed from the effective rank. The released rank budget is obtained by subtracting the effective rank value after rank pruning from the effective rank value before rank pruning.

[0119] The release rank budget is sorted and redistributed based on marginal benefits. Marginal benefits are calculated on the validation data. The marginal benefit calculation process includes selecting each adaptation layer identifier in the adaptation layer set, increasing the effective rank corresponding to the adaptation layer identifier by one unit to obtain the candidate effective rank, updating the hierarchical rank allocation table based on the candidate effective rank and performing validation inference to calculate the candidate target index. The change in the target index is obtained by subtracting the baseline target index from the candidate target index. The marginal benefit is the negative of the change in the target index. The adaptation layer identifiers are sorted based on the marginal benefits to obtain the marginal benefit sorting result.

[0120] The release rank budget is reallocated based on the marginal revenue ranking results. The reallocation process selects the adaptation layer identifier according to the order of the marginal revenue ranking results. In the hierarchical rank allocation table, the effective rank corresponding to the selected adaptation layer identifier is updated by one unit. This process is repeated until the release rank budget is allocated. After the allocation is completed, the updated hierarchical rank allocation table is output.

[0121] In this embodiment, the output of the fusion result specifically includes:

[0122] Read the final hierarchical rank allocation table and perform conditional activation on the adaptation layer set. The conditional activation process traverses each adaptation layer identifier in the adaptation layer set and reads the effective rank value corresponding to the adaptation layer identifier in the final hierarchical rank allocation table. The effective rank value is used to determine the number of rank components in the low-rank incremental parameter structure participating in the inference calculation.

[0123] The large model set is invoked to perform inference on the input data. During the inference process, forward computation is performed on each large model in the large model set in sequence. During the forward computation, the low-rank incremental parameter structure determined by conditional activation is used at the adaptation layer set to participate in the computation. The output model output is associated with the model output model identifier.

[0124] For each large model in the large model set, the marginal benefit value and the conflict marker count value are read. The model score calculation process receives the marginal benefit value and the conflict marker count value as input, and maps the marginal benefit value and the conflict marker count value to a non-negative model score value according to the preset calculation rules. The preset calculation rules are set and kept fixed before the end of training. The fusion coefficient calculation process performs normalization processing on all model score values ​​in the large model set. The normalization processing divides each model score value by the sum of the model score values ​​to obtain the fusion coefficient, so that the sum of the fusion coefficients in the dimension of the large model set is equal to 1.

[0125] Adaptive fusion is performed on the model output according to the fusion coefficient. The adaptive fusion process multiplies the output of each model in the large model set by the corresponding fusion coefficient to obtain the weighted model output. The weighted model output is then summed to obtain the fusion result, which is then output.

[0126] A large-model adaptive fusion system based on deep learning includes:

[0127] The data and task acquisition module is used to acquire input data and task identifiers;

[0128] The configuration module is used to determine the large model set and the adaptation layer set based on the task identifier, and to establish a low-rank incremental parameter structure for each adaptation layer in the adaptation layer set.

[0129] The hierarchical rank allocation table generation module is used to encode the input data to obtain the context vector, generate the model-level rank budget and the hierarchical rank upper limit, and generate the hierarchical rank allocation table based on the model-level rank budget, the hierarchical rank upper limit, and the total rank budget.

[0130] The training update module is used to perform improved AdaLoRA training updates based on the hierarchical rank allocation table, train the low-rank incremental parameter structure and dynamically adjust the effective rank.

[0131] The meta-update module is used to calculate the target index based on the validation data, perform meta-updates on the model-level rank budget and the hierarchical rank upper limit based on the target index, and update the hierarchical rank allocation table based on the meta-update results.

[0132] The conflict marking module is used to convert the low-rank incremental parameter structure into a subspace representation, calculate the distance between subspaces, and generate conflict markers based on the distance between subspaces.

[0133] The rank redistribution module is used to perform orthogonal complementary projection on the subspace representation corresponding to the conflict marker, perform rank pruning to obtain the release rank budget, sort and redistribute the release rank budget based on marginal benefits, and update the hierarchical rank allocation table.

[0134] The fusion inference module is used to perform conditional activation based on the final hierarchical rank allocation table, generate fusion coefficients based on marginal benefit values ​​and conflict marker counts, and perform adaptive fusion on the output of the large model set based on the fusion coefficients to output the fusion result.

[0135] Example 1:

[0136] To verify the feasibility of this invention in practice, it was applied to a large-scale model service system for enterprise knowledge-based question answering and automated work order processing. This system simultaneously undertakes four tasks: knowledge base retrieval and question answering, work order classification and routing, fault cause summarization, and processing suggestion generation. Input data includes user natural language descriptions, structured work order fields, and historical dialogue context. Task identifiers are automatically provided by the business entry point. Existing solutions use a single large model for unified service, with full-scale fine-tuning or fixed-rank LoRA fine-tuning performed on a small number of business processes. This leads to two prominent problems: First, different tasks have significantly different requirements for model capabilities, and fixed-rank configurations cause high-frequency tasks to consume a large amount of parameter budget while low-frequency tasks experience decreased performance. Second, when multiple models are fused in parallel, the incremental parameters of the adaptation layer learn similar directions, increasing output divergence and causing unstable fusion weights. Online performance exhibits phenomena such as fluctuating conclusions, duplicate generated content, and conflicting details for the same input across different models, resulting in work order routing errors and decreased credibility of knowledge-based question answering.

[0137] In this scenario, the system is configured with a large model set containing three basic models. The adaptation layer set selects attention projection linear layers and feedforward network linear layers, and establishes a low-rank incremental parameter structure for each adaptation layer. After receiving input data and task identifiers, the system encodes the input to obtain a context vector. The context vector drives the generation of model-level rank budget and hierarchical rank upper limit, and generates a hierarchical rank allocation table under the constraint of the total rank budget, so that each model obtains a clear effective rank value at each adaptation layer. During the training phase, an improved AdaLoRA training update is used. Parameter updates are performed on the rank components corresponding to the effective rank indicated by the hierarchical rank allocation table, and freezing control is performed on unindicated rank components. At the same time, within a preset training period, the rank importance score is calculated based on the training loss. Rank pruning and rank addition are performed according to the first and second thresholds, and the change magnitude constraint is applied to stabilize the evolution of effective rank. During the validation phase, validation output is obtained through validation data inference. The validation loss is calculated and combined with the total rank budget penalty term to form a target index. The target index is used to evaluate the model. The rank budget and the upper limit of the hierarchical rank are updated meta-updated, thereby updating the hierarchical rank allocation table. This allows resource allocation to be automatically adjusted according to changes in task difficulty and input distribution. Subsequently, the low-rank incremental parameter structure is converted into a subspace representation. The distance between subspaces of the adaptation layers is calculated and conflict markers are generated. Orthogonal complementary projection and rank pruning are performed on the subspace representations corresponding to the conflict markers to release the rank budget. Then, rank redistribution is performed based on marginal revenue as the sorting criterion, and the hierarchical rank allocation table is updated. During the inference phase, the final hierarchical rank allocation table is read to perform conditional activation on the adaptation layer set. The marginal revenue value and conflict marker count value are converted into model scores according to preset calculation rules, normalized to obtain fusion coefficients, and adaptive fusion is performed on the output of the large model set to obtain the final result. Thus, stable fusion and task adaptation are achieved without expanding the total rank budget.

[0138] To verify the beneficial effects, real business data from the system were selected to construct training and validation sets. The training set contained 1,200,000 samples, and the validation set contained 80,000 samples, covering four types of tasks and maintaining consistency with the online task distribution. Work order classification routing was evaluated using macro-average F1 and routing accuracy; knowledge question answering was evaluated using answer consistency scoring and citation hit rate; cause summarization was evaluated using manual review pass rate; and handling suggestions were evaluated using standard clause hit rate and duplication rate. Simultaneously, engineering indicators such as parameter overhead, inference throughput, and latency were recorded. Comparison solutions included fixed... This invention comprises a simplified version that includes a fixed-rank LoRA single model, a three-model average fusion of fixed-rank LoRA, a three-model fusion with dynamic routing gating, and a version without conflict marking and rank reassignment. Under the condition of a total rank budget of 512, the final hierarchical rank allocation table maintains an effective total rank of 512 after training convergence, exhibiting task-driven model-level rank budget differences. The model budgets for work order classification and routing tasks are improved, while the upper limit of hierarchical rank for knowledge question answering tasks is increased on the attention projection linear layer. The upper limit of the feedforward network linear layer remains low to reduce redundancy. The verification results are shown in Table 1.

[0139] Table 1. Comparison of Fusion and Fine-tuning Effects under Resource-Constrained Conditions

[0140] plan General budget Number of compatible parameters (M) Work order routing accuracy (%) Work order classification F1 (%) Knowledge Q&A Citation Hit Rate (%) Reasons for review and approval rate (%) Repeat rate of treatment recommendations (%) Inference throughput (messages / second) P95 latency (ms) Single model + fixed-rank LoRA 512 18.6 90.8 88.9 71.4 83.6 14.2 22.1 980 Three-model average fusion + fixed-rank LoRA 512×3 55.8 92.1 90.3 74.9 85.2 13.6 14.7 1460 Three-model dynamic routing gating fusion 512×3 55.8 92.4 90.7 75.6 85.7 13.1 16.2 1320 This invention is a simplified version (without conflict markers and rank allocation). 512 18.6 93.0 91.5 77.2 87.1 11.8 20.4 1060 This invention 512 18.6 94.3 93.1 80.8 90.4 9.2 20.1 1090

[0141] As the data shows, with the total rank budget maintained at 512 and the number of adaptation parameters maintained at 18.6M, this invention improves the accuracy of work order routing, classification F1, citation hit rate, and review pass rate, and reduces the repetition rate of handling suggestions from 14.2% to 9.2%. Compared with the average fusion of three models and dynamic routing gating fusion, this invention avoids configuring a total rank budget of 512 independently for each model, reduces the number of adaptation parameters by 37.2M, and maintains inference throughput at the level of 20 messages / second, with P95 latency controlled at 1090ms. Compared with the abridged version, after introducing subspace conflict marking, orthogonal complementary projection, and rank pruning to release the rank budget and redistributing it according to marginal benefits, the citation hit rate increases by 3.6 percentage points, the review pass rate increases by 3.3 percentage points, and the repetition rate decreases by 2.6 percentage points, reflecting that conflict suppression and resource redistribution have a direct effect on fusion consistency and generation quality.

[0142] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A large model self-adaptive fusion method based on deep learning, characterized in that, The method comprises the following steps: obtaining input data and task identification, determining a large model set and an adaptive layer set, and establishing a low-rank incremental parameter structure for each adaptive layer in the adaptive layer set; encoding the input data to obtain a context vector, generating a model-level rank budget and a hierarchical rank upper limit based on the context vector, and generating a hierarchical rank allocation table based on the model-level rank budget, the hierarchical rank upper limit, and a total rank budget; performing improved AdaLoRA training and updating based on the hierarchical rank allocation table, training the low-rank incremental parameter structure, and dynamically adjusting the effective rank; calculating a target index based on validation data, performing meta-updating on the model-level rank budget and the hierarchical rank upper limit based on the target index, and updating the hierarchical rank allocation table based on the meta-updating result; converting the low-rank incremental parameter structure into a subspace representation, calculating the distance between subspaces, and generating a conflict marker based on the distance between subspaces; performing orthogonal complementary projection on the subspace representation corresponding to the conflict marker, performing rank clipping to obtain a released rank budget, and re-distributing and updating the hierarchical rank allocation table based on the marginal benefit of the released rank budget; outputting a final hierarchical rank allocation table and an effective rank based on a preset termination condition; in the inference stage, performing conditional activation based on the final hierarchical rank allocation table, generating a fusion coefficient based on the marginal benefit and the conflict marker, performing adaptive fusion on the large model set output based on the fusion coefficient, and outputting a fusion result.

2. The large model self-adaptive fusion method based on deep learning according to claim 1, characterized in that, The determination of the large model set and the adaptive layer set and the establishment of the low-rank incremental parameter structure for each adaptive layer in the adaptive layer set specifically comprises: determining the large model set and the adaptive layer set based on the task identification, the large model set containing at least two large models, each large model corresponding to a model identification; for each large model in the large model set, filtering to obtain the adaptive layer set based on a preset layer type set, the preset layer type set containing attention projection linear layers and feedforward network linear layers; establishing a low-rank incremental parameter structure for each adaptive layer in the adaptive layer set, the low-rank incremental parameter structure containing a first low-rank parameter matrix and a second low-rank parameter matrix, the connection dimension of the first low-rank parameter matrix and the second low-rank parameter matrix serving as the effective rank; constructing an updated weight matrix based on the adaptive layer weight matrix, the low-rank incremental parameter structure, and a scaling coefficient, the scaling coefficient being set before the start of training and corresponding to the adaptive layer identification one by one.

3. The large model self-adaptive fusion method based on deep learning according to claim 1, characterized in that, The generation of the hierarchical rank allocation table specifically comprises: encoding the input data to obtain a context vector; generating a model-level rank budget and a hierarchical rank upper limit based on the context vector, the model-level rank budget corresponding to the large model set one by one in the dimension, and the hierarchical rank upper limit corresponding to the adaptive layer set one by one in the dimension; constructing a hierarchical rank allocation table under the constraint of a total rank budget, the hierarchical rank allocation table being a two-dimensional index structure composed of large model identifications and adaptive layer identifications, each index position corresponding to an effective rank value; performing allocation on the hierarchical rank allocation table according to the large model dimension under the constraint of the model-level rank budget, so that the sum of the effective ranks corresponding to each large model is equal to the model-level rank budget; performing clipping on the hierarchical rank allocation table according to the adaptive layer dimension under the constraint of the hierarchical rank upper limit, so that the effective rank corresponding to each adaptive layer does not exceed the hierarchical rank upper limit; Perform global correction on the allocated and pruned hierarchical rank allocation table, so that the sum of all valid ranks in the hierarchical rank allocation table is equal to the total rank budget, and complete the hierarchical rank allocation table generation.

4. The large model self-adaptive fusion method based on deep learning according to claim 1, characterized in that, The definition and training update of the improved AdaLoRA specifically include: Based on the hierarchical rank allocation table, determine the valid rank corresponding to each adaptive layer in the adaptive layer set, and based on the valid rank, impose a structural constraint on the low-rank incremental parameter structure, so that the number of rank components participating in training in the low-rank incremental parameter structure is consistent with the valid rank; Based on the low-rank incremental parameter structure, an improved AdaLoRA training mechanism is used to perform training updates on the adaptive layer set, and the improved AdaLoRA training mechanism includes: During the training process, only the parameters of the rank components corresponding to the valid ranks indicated by the hierarchical rank allocation table are updated, and the rank components not indicated by the hierarchical rank allocation table are controlled to be frozen, so that the frozen rank components keep their parameters unchanged during the training process; Within a preset training period, based on the training loss, calculate the rank importance score of each rank component in the low-rank incremental parameter structure, and the rank importance score is determined based on the gradient amplitude and parameter amplitude of the corresponding rank component; Based on the rank importance score and the hierarchical rank allocation table, perform rank evolution control, and when the rank importance score is lower than a first threshold, perform a pruning operation on the corresponding rank component, and when the rank importance score is higher than a second threshold and there is available rank budget in the hierarchical rank allocation table, perform an adding operation on the corresponding rank component; During the rank evolution control process, a change amplitude constraint is imposed on the pruning operation and the adding operation, so that the number of valid rank changes of the adaptive layer within a single training period does not exceed a preset upper limit of change, and the valid rank in the low-rank incremental parameter structure is updated synchronously after the change is completed.

5. The large model self-adaptive fusion method based on deep learning according to claim 1, characterized in that, The updating of the hierarchical rank allocation table based on the meta-update result specifically includes: Based on the validation data, construct a validation batch, perform conditional activation on the adaptive layer set based on the hierarchical rank allocation table, and perform validation inference on the large model set after the conditional activation to obtain a validation output. Based on the validation output and the validation label, calculate a validation loss, which is the function value of the validation output and the validation label under a preset loss function. Based on the validation loss, calculate a target indicator, which is obtained by combining the validation loss and a total rank budget penalty term determined by the difference between the sum of all valid ranks in the hierarchical rank allocation table and the total rank budget, and the target indicator is one-to-one corresponding to the task identifier. Based on the target indicator, perform meta-updating on the model-level rank budget and the hierarchical rank upper limit, which includes calculating the sensitivity of the target indicator to the model-level rank budget and the hierarchical rank upper limit and generating an update increment, and updating the model-level rank budget and the hierarchical rank upper limit based on the update increment to obtain an updated model-level rank budget and an updated hierarchical rank upper limit. Based on the updated model-level rank budget, the updated hierarchical rank upper limit, and the total rank budget, update the hierarchical rank allocation table, so that the sum of all valid ranks in the hierarchical rank allocation table is equal to the total rank budget, the valid rank corresponding to each adaptive layer in the hierarchical rank allocation table does not exceed the updated hierarchical rank upper limit, and the sum of the valid ranks corresponding to each large model in the hierarchical rank allocation table is equal to the updated model-level rank budget.

6. The large model self-adaptive fusion method based on deep learning according to claim 1, characterized in that, The generation of the conflict marker specifically includes: The low-rank incremental parameter structure is traversed based on the adaptation layer set, and for each adaptation layer in the adaptation layer set, the low-rank incremental parameter structure is converted into a subspace representation; The subspace representation is determined by a first low-rank parameter matrix and a second low-rank parameter matrix, and the subspace representation includes effective rank basis vectors, and the basis vectors are determined by column vectors of the first low-rank parameter matrix, and the basis vectors correspond to the effective rank one by one; A subspace pair set is constructed in the adaptation layer set dimension, and the subspace pair set includes an adaptation layer identifier pair and a subspace representation pair, the adaptation layer identifier pair is composed of two different adaptation layer identifiers in the adaptation layer set, and the subspace representation pair is composed of two subspace representations corresponding to the adaptation layer identifier pair; The subspace distance is calculated for each subspace pair, and the subspace distance is obtained by performing distance measurement operation on the basis vector set of the subspace representation pair; A conflict marker is generated based on the subspace distance, and the conflict marker is generated for the adaptation layer identifier pair when the subspace distance is less than a distance threshold, and the conflict marker is not generated when the subspace distance is greater than the distance threshold, and the distance threshold is set before the training starts and corresponds to the task identifier one by one.

7. The large model self-adaptive fusion method based on deep learning according to claim 1, characterized in that, The release rank budget is sorted and redistributed based on the marginal benefit, and the hierarchical rank allocation table is updated, which specifically includes: The adaptation layer identifier pair set is determined based on the conflict marker, and the adaptation layer identifier pair set is composed of the adaptation layer identifier pairs corresponding to the conflict markers; For each adaptation layer identifier pair in the adaptation layer identifier pair set, the subspace representation pair corresponding to the adaptation layer identifier pair is read, and for each subspace representation in the subspace representation pair, an orthogonal complement projection is performed, the orthogonal complement projection projects the subspace representation onto the orthogonal complement space of the other subspace representation in the subspace representation pair, and the projected subspace representation is outputted; Rank pruning is performed based on the projected subspace representation, and the rank pruning includes determining a rank component set participating in training in the projected subspace representation, and removing the rank components not entering the rank component set from the effective rank, and the release rank budget is determined based on the difference between the effective rank before the rank pruning and the effective rank after the rank pruning; The release rank budget is sorted and redistributed based on the marginal benefit, and the marginal benefit is calculated on the validation data, and the marginal benefit is the opposite of the target index change amount when the effective rank of the adaptation layer identifier pair in the adaptation layer set is increased by one unit, and the target index change amount is determined by the difference between the target index before the effective rank is increased by one unit and the target index after the effective rank is increased by one unit; According to the marginal benefit sorting result, the release rank budget is redistributed, and the redistribution includes allocating the effective rank increment to the adaptation layer identifier in the adaptation layer set in the order from large to small according to the marginal benefit, and the hierarchical rank allocation table is updated based on the allocation result.

8. The large model self-adaptive fusion method based on deep learning according to claim 1, characterized in that, The output of the fusion result specifically includes: The final hierarchical rank allocation table is read and the conditional activation is performed on the adaptation layer set; The inference is performed on the input data by calling the large model set to obtain the model output corresponding to each large model in the large model set; For each large model in the large model set, the marginal benefit value and the conflict marker count value are read, the marginal benefit value and the conflict marker count value are converted into a model score according to a pre-designed calculation rule, and the fusion coefficient is obtained by normalizing the model score in the large model set dimension. Adaptive fusion is performed on the model output according to the fusion coefficient, and a fusion result is output.

9. A deep learning-based large model self-adaptive fusion system, which executes the deep learning-based large model self-adaptive fusion method according to any one of claims 1 to 8. Comprise: A data and task acquisition module for acquiring input data and a task identifier; A configuration module for determining a large model set and an adaptation layer set according to the task identifier, and establishing a low-rank incremental parameter structure for each adaptation layer in the adaptation layer set; A hierarchical rank allocation table generation module for encoding the input data to obtain a context vector, generating a model-level rank budget and a hierarchical rank upper limit, and generating a hierarchical rank allocation table based on the model-level rank budget, the hierarchical rank upper limit, and a total rank budget; A training update module for performing improved AdaLoRA training updates based on the hierarchical rank allocation table, training the low-rank incremental parameter structure, and dynamically adjusting the effective rank; An update module for calculating a target index based on validation data, performing meta-updates on the model-level rank budget and the hierarchical rank upper limit based on the target index, and updating the hierarchical rank allocation table based on the meta-update result; A conflict marking module for converting the low-rank incremental parameter structure into a subspace representation, calculating the distance between subspaces, and generating a conflict marker based on the distance between subspaces; A rank reallocation module for performing orthogonal complement projection on the subspace representation corresponding to the conflict marker, performing rank clipping to obtain a released rank budget, sorting and reallocating the released rank budget based on marginal benefits, and updating the hierarchical rank allocation table; A fusion inference module for performing conditional activation based on the final hierarchical rank allocation table, generating a fusion coefficient based on the marginal benefit value and the conflict marker count value, and performing adaptive fusion on the large model set output based on the fusion coefficient to output a fusion result.

Images