Large model adjustment method and device, equipment and storage medium

Abstract

The application relates to the technical field of large models, and discloses a large model adjusting method and device, equipment and a storage medium. The method comprises the following steps: performing segmented deployment on a large model to be adjusted to obtain sub-models deployed on different terminal side devices, wherein each sub-model comprises a different number of coding and decoding layers and an adjustable adapter, and each coding and decoding layer is embedded with a corresponding adjustable adapter; updating the adjustable adapters layer by layer according to the training state of the large model to be adjusted to obtain a large model embedded with updated configuration of the adjustable adapters; and performing forward propagation calculation and backward propagation calculation on the large model embedded with updated configuration of the adjustable adapters to obtain a large model after parameter adjustment. The application can effectively improve the large model parameter adjustment efficiency and save device resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large model, in particular to a large model adjusting method and device, equipment and storage medium. BACKGROUND

[0002] In recent years, general-purpose large models based on the Transformer architecture have made significant progress in the field of artificial intelligence and have shown excellent performance in natural language processing, computer vision, and other tasks. These large models, through pre-training on large-scale datasets, have acquired strong generalization representation capabilities and, in the fine-tuning stage, have adapted to specific tasks, achieving excellent performance in multiple benchmark tests. The fine-tuning stage of large models relies on sample data in specific domains, which are widely distributed on end-side devices (such as smartphones, Internet of Things devices, etc.) and continuously generated in the process of user interaction responses. In order to effectively utilize distributed data while protecting data privacy, using end-side devices to perform large model fine-tuning intelligence tasks has become a current hot technology. Mainstream general-purpose large models usually have multiple Transformer modules, each containing multi-head self-attention and feed-forward neural networks. This hierarchical structure enables the Transformer to efficiently learn complex data patterns, but also introduces large-scale matrix operations. Executing such high-computational intelligence tasks on end-side devices faces significant computational resource bottlenecks, easily leading to processing delays and excessive energy consumption. In addition, the parameter quantity of large models is usually in the hundreds of millions or even billions, and a large amount of memory is needed to store intermediate results such as activation values, gradients, and optimizer states during forward / backward propagation calculations, posing a huge challenge to the memory capacity of end-side devices, making it prone to memory overflow problems during model loading and data processing, causing fine-tuning tasks to be interrupted abnormally. By exploiting the communication interconnection and resource complementation capabilities between network-connected end-side devices, using model segmentation and multi-device pipeline parallel computing technology, large models can be divided into multiple sub-models and distributed to different end-side devices for collaborative execution, which can significantly reduce the computational and memory resource consumption of a single device. These end-side devices coexist in a local area network and form a pipeline topology according to the hierarchical order of the sub-models they deploy in the original large model. The data samples required for fine-tuning are stored in the "head" device of the pipeline, and the corresponding sample labels are passed to the "tail" device of the pipeline. In the forward propagation process, each device applies the sub-model layer it is responsible for to the input passed from the previous stage and passes the output of the last layer to the next stage device. Backward propagation is performed in reverse order, and gradient information is passed back to each device in turn. To improve training efficiency, multiple devices can process different batches of data simultaneously, and each device saves a copy of the sub-model weights used in forward propagation to facilitate subsequent backward propagation for the corresponding batch.

[0003] However, the existing pipeline parallel-based cooperative fine-tuning technology mainly has the following two problems, 1) the current pipeline parallel-based cooperative training design generally follows a chain sequence training mechanism, which requires the first device in the chain to hold all input data and share the relevant labels with the last device, which raises privacy issues and increases interaction overhead. In addition, only the data samples of the "head" device of the pipeline are allowed to train during the entire process, while other devices only serve as auxiliary computing nodes. Therefore, this pipeline parallel training framework cannot utilize the data of multiple devices to fine-tune a unified model. In practical applications, due to storage limitations, the data samples on a single device may be quite limited, while multiple devices within a local area network can often collect different types of task data, and the inefficient use of these data will compromise the performance of the fine-tuned model. 2) Pipeline parallel training allows forward propagation calculations using un-updated model parameters, and to update the direction of the unified model, each device is required to store different versions of model parameters for average updating, which will introduce significant storage overhead, making the memory and storage capacity of the edge devices already insufficient to face more severe resource challenges, which can easily lead to the failure of the fine-tuning task.

[0004] Therefore, how to effectively improve the large model tuning efficiency and save device resources is a problem that needs to be solved at present.

[0005] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0006] The main purpose of the present application is to provide a large model adjustment method, device and storage medium, which aims to solve the technical problem of how to effectively improve the large model tuning efficiency and save device resources.

[0007] To achieve the above purpose, the present application provides a large model adjustment method, which comprises:

[0008] Segmenting and deploying the large model to be adjusted to obtain sub-models deployed on different edge devices, wherein each sub-model comprises a different number of encoding and decoding layers and an adjustable adapter, and each encoding and decoding layer is embedded with a corresponding adjustable adapter;

[0009] Updating the adjustable adapters layer by layer according to the training state of the large model to be adjusted to obtain a large model with updated embedded configuration adapters;

[0010] Performing forward propagation calculation and backward propagation calculation on the large model with updated embedded configuration adapters to obtain a large model after parameter adjustment.

[0011] In an embodiment, the segmenting and deploying the large model to be adjusted to obtain multiple sub-models and corresponding edge devices comprises:

[0012] determining a number of codec layers for each end-side device according to a memory capacity of each end-side device;

[0013] segmenting the large model to be adjusted into a plurality of sub-models according to the number of codec layers;

[0014] obtaining a target Hamilton circuit;

[0015] deploying the plurality of sub-models in the corresponding end-side devices in sequence based on the target Hamilton circuit to obtain sub-models deployed in different end-side devices.

[0016] In an embodiment, the determining a number of codec layers for each end-side device according to a memory capacity of each end-side device comprises:

[0017] obtaining a plurality of end-side devices and corresponding forward propagation calculation delays, and sorting the end-side devices based on the forward propagation calculation delays to obtain sorted end-side devices;

[0018] determining a number of codec layers for each end-side device according to a memory capacity of each end-side device;

[0019] determining a number of codec layers for each end-side device according to the number of codec layers and the sorted end-side devices.

[0020] In an embodiment, the obtaining a target Hamilton circuit comprises:

[0021] obtaining transmission links between each end-side device, and constructing a directed graph according to each end-side device and the transmission links, wherein the directed graph comprises a plurality of original nodes;

[0022] adding virtual nodes in the directed graph, and constructing directed edges based on the virtual nodes and each original node;

[0023] obtaining weights of each directed edge, and generating a directed weighted graph based on the weights of each directed edge;

[0024] determining a target Hamilton circuit based on the directed weighted graph.

[0025] In an embodiment, the determining a target Hamilton circuit based on the directed weighted graph comprises:

[0026] taking the virtual nodes in the directed weighted graph as starting points and performing path planning to generate a path table;

[0027] determining path lengths from each visited node in a set of visited nodes to unvisited nodes based on the path table;

[0028] determine whether a reference Hamiltonian loop exists according to the path length;

[0029] when the reference Hamiltonian loop exists, removing the virtual node and reconstructing a path through a backtracking table to generate a target Hamiltonian loop.

[0030] In an embodiment, the sequentially deploying a plurality of the sub-models in corresponding end-side devices based on the target Hamiltonian loop obtains sub-models deployed in different end-side devices, including:

[0031] determining an end-side device arrangement order based on the target Hamiltonian loop;

[0032] obtaining model parameters of each sub-model;

[0033] sequentially sending the model parameters to corresponding end-side devices according to the end-side device arrangement order to obtain sub-models deployed in different end-side devices.

[0034] In an embodiment, the updating the adjustable adapter layer by layer according to the training state of the to-be-adjusted large model to obtain a large model embedding an updated configuration of the adjustable adapter, including:

[0035] obtaining a connection order of a coding and decoding layer in each sub-model in the to-be-adjusted large model;

[0036] activating a corresponding adjustable adapter layer by layer according to the training state of the to-be-adjusted large model and the connection order to obtain an activated state of the adjustable adapter;

[0037] adjusting a width configuration of the activated state of the adjustable adapter to obtain an updated configuration of the adjustable adapter;

[0038] embedding the updated configuration of the adjustable adapter into the to-be-adjusted large model to obtain a large model embedding an updated configuration of the adjustable adapter.

[0039] In an embodiment, the adjusting the width configuration of the activated state of the adjustable adapter to obtain an updated configuration of the adjustable adapter, including:

[0040] obtaining gradient information of a current adjustable parameter of the activated state of the adjustable adapter;

[0041] determining a trace of a Fisher information matrix based on the gradient information of the current adjustable parameter;

[0042] sorting the activated state of the adjustable adapter based on the trace of the Fisher information matrix to obtain a target adapter;

[0043] performing bottleneck dimension updating on the target adapter based on a preset bottleneck dimension growth step to obtain an updated configuration of the adjustable adapter.

[0044] In an embodiment, the forward propagation calculation and the backward propagation calculation are performed on the large model in which the updated configurable adapter is embedded, to obtain a large model with adjusted parameters.

[0045] The forward propagation calculation and the backward propagation calculation are performed on the sub-models of the end-side devices based on the task data samples, to implement pipeline parallel training, and obtain a trained large model.

[0046] The training convergence is evaluated according to the loss value of the trained large model, to obtain an evaluation result.

[0047] When the evaluation result is that the training is converged, the trained large model is taken as the large model with adjusted parameters.

[0048] When the evaluation result is that the training is not converged, the trained large model is subjected to polling training until the training is converged, to obtain the large model with adjusted parameters.

[0049] In an embodiment, the forward propagation calculation and the backward propagation calculation are performed on the sub-models of the end-side devices based on the task data samples, to implement pipeline parallel training, and obtain a trained large model, including:

[0050] The forward propagation path and the backward propagation path are determined based on the end-side devices.

[0051] The sub-models of the end-side devices are subjected to forward propagation training along the forward propagation path and to backward propagation training along the backward propagation path based on the task data samples, to implement pipeline parallel training, and obtain a trained large model.

[0052] In addition, to achieve the above object, the present application further provides a large model adjustment device, which comprises:

[0053] A deployment module is configured to perform segmented deployment on a large model to be adjusted, to obtain sub-models deployed on different end-side devices, wherein each sub-model comprises a different number of codec layers and adjustable adapters, and each codec layer is embedded with a corresponding adjustable adapter.

[0054] An updating module is configured to update the adjustable adapters layer by layer according to a training state of the large model to be adjusted, to obtain a large model in which the updated configurable adapters are embedded.

[0055] A propagation module is configured to perform forward propagation calculation and backward propagation calculation on the large model in which the updated configurable adapters are embedded, to obtain a large model with adjusted parameters.

[0056] In addition, to achieve the above-mentioned purpose, the present application also proposes a large model adjustment device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program is configured to implement the steps of the large model adjustment method as described above.

[0057] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium. A computer program is stored on the storage medium, and when the computer program is executed by the processor, the steps of the large model adjustment method described above are implemented.

[0058] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the large model adjustment method described above.

[0059] The present application provides a large model adjustment method. The present application first divides and deploys the large model to be adjusted to obtain sub-models deployed on different end-side devices, wherein each sub-model includes a different number of codec layers and adjustable adapters, and a corresponding adjustable adapter is embedded after each codec layer; the adjustable adapter is updated layer by layer according to the training status of the large model to be adjusted to obtain a large model with the adjustable adapter embedded in the updated configuration; forward propagation calculation and backward propagation calculation are performed on the large model with the adjustable adapter embedded in the updated configuration to obtain a large model with adjusted parameters, which can effectively improve the efficiency of large model parameter adjustment and save device resources.

[0060] In summary, this application can accurately adapt to the corresponding end-side devices by splitting and deploying the large model to be adjusted, obtain sub-models deployed on different end-side devices, and update the adjustable adapters in each sub-model layer by layer, effectively shortening the backpropagation path, thereby performing forward propagation and backward propagation, and realizing fine-tuning of the large model through multiple devices, and quickly and efficiently obtaining the adjusted large model, overcoming the technical defects of being unable to use multiple devices for model fine-tuning and the failure of model fine-tuning due to limited device resources, and can effectively improve the efficiency of large model parameter adjustment and save device resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0062] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0063] Figure 1 A flow chart of the first embodiment of the large model adjustment method of the present application is provided;

[0064] Figure 2 A schematic diagram of pipeline parallel training provided in Example 1 of the large model adjustment method of this application;

[0065] Figure 3 This is a flowchart of the online configuration decision-making process of the adapter provided in Example 1 of the large model adjustment method of this application;

[0066] Figure 4 A schematic diagram of the Transformer large model structure of the integrated serial adapter provided in Example 1 of the large model adjustment method of this application;

[0067] Figure 5 A schematic diagram of multi-device collaborative fine-tuning training provided in Example 1 of the large model adjustment method of this application;

[0068] Figure 6 A flow chart of the second embodiment of the large model adjustment method of this application is provided;

[0069] Figure 7 A model segmentation and deployment decision flow chart provided in Example 2 of the large model adjustment method of this application;

[0070] Figure 8 This is a schematic diagram of the module structure of the large model adjustment device according to an embodiment of the present application;

[0071] Figure 9 This is a schematic diagram of the device structure of the hardware operating environment involved in the large model adjustment method of the embodiment of the present application.

[0072] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0073] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0074] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0075] The main solution of the embodiment of the present application is: the to-be-adjusted large model is segmented and deployed to obtain sub-models deployed on different end-side devices, wherein each sub-model includes different numbers of codec layers and adjustable adapters, and each codec layer is embedded with a corresponding adjustable adapter; the adjustable adapters are updated layer by layer according to the training state of the to-be-adjusted large model to obtain a large model embedded with updated configuration of the adjustable adapters; and the large model embedded with updated configuration of the adjustable adapters is subjected to forward propagation calculation and backward propagation calculation to obtain a large model after parameter adjustment.

[0076] The existing pipeline parallel-based collaborative fine-tuning technology mainly has the following two problems, 1) the current pipeline parallel-based collaborative training design generally follows a chain sequence training mechanism, which requires the first device in the chain to hold all input data and share the related labels with the last device, which raises privacy problems and increases interaction overhead. In addition, only the data samples of the "head" device of the pipeline are allowed to be trained in the whole process, while other devices only serve as auxiliary computing nodes. Therefore, this pipeline parallel training framework cannot utilize the data of multiple devices to fine-tune a unified model. In actual applications, due to storage limitations, the data samples on a single device can be quite limited, while multiple devices in a local area network can often collect different types of task data, and the inefficient use of these data will harm the performance of the fine-tuned model. 2) The pipeline parallel training allows forward propagation calculation using un-updated model parameters, and to update the direction of the unified model, each device is required to store different versions of model parameters for average update, which will introduce significant storage overhead, making the end-side devices with insufficient memory and storage capacity face more severe resource challenges, which easily leads to the failure of the fine-tuning task. Therefore, how to effectively improve the large model parameter adjustment efficiency and save device resources is a problem that needs to be solved at present.

[0077] The present application can accurately adapt to the corresponding end-side device by segmenting and deploying the to-be-adjusted large model, obtain sub-models deployed on different end-side devices, update the adjustable adapters in each sub-model layer by layer, effectively shorten the backpropagation path, and then perform forward propagation and backward propagation, thereby fine-tuning the large model through multiple devices and quickly and efficiently obtaining the adjusted large model. The technical defects of being unable to utilize multiple devices for model fine-tuning and the limited device resources leading to the failure of model fine-tuning are overcome, the large model parameter adjustment efficiency is effectively improved, and device resources are saved.

[0078] It should be noted that the execution subject of the embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a large model adjustment device, etc. The embodiments will be described below with the large model adjustment device as an example.

[0079] Based on this, the large model adjustment method provided in the embodiments of the present application comprises the following steps: Figure 1 , Figure 1 The flowchart of the first embodiment of the large model adjustment method of the present application. In this embodiment, the large model adjustment method comprises steps S10-S30:

[0080] Step S10, segmenting and deploying the large model to be adjusted to obtain sub-models deployed on different end-side devices, wherein each sub-model comprises a different number of encoding and decoding layers and an adjustable adapter, and each encoding and decoding layer is embedded with a corresponding adjustable adapter.

[0081] It should be noted that a large model is a machine learning model containing a large number of parameters, such as a deep learning network, which is usually used to process complex tasks such as natural language processing, image recognition, etc. In this embodiment, the large model to be adjusted is a pre-trained model that already has the basic ability to handle specific tasks, but in order to adapt to specific application scenarios, further adjustment or fine-tuning is needed, for example, in the distributed computing power network of smart home, etc., supporting a number of power and memory limited network-side devices to cooperate with each other to fine-tune the complex Transformer large model training using their local label data.

[0082] It can be understood that the large model to be adjusted comprises multiple layers of encoding and decoding layers, and the encoding and decoding layer, i.e. the transforme layer, is a coding and decoding structure used to process the conversion between input data and output data. In the embodiments, the encoding and decoding layers of the large model to be adjusted are segmented and deployed on different end-side devices, and each device deployed with a sub-model containing a different number of encoding and decoding layers to adapt to the computing power and storage resources of the device.

[0083] It is worth noting that the adjustable adapter is a mechanism for adjusting model parameters, which can fine-tune the model according to the specific needs and environmental conditions of the end-side device. In this embodiment, each encoding and decoding layer is embedded with a corresponding adjustable adapter, so that each sub-model can independently adjust the parameters to adapt to different application scenarios and data distribution.

[0084] It is worth noting that splitting the large model to be adjusted across a group of end-side devices can adapt to the differentiated computing speeds, memory capacities, and link rates of these devices, addressing the computing power and memory resource limitations of a single device supporting large model fine-tuning tasks. An adjustable adapter layer is embedded after each Transformer layer carried by each end-side device. During fine-tuning, only the parameters of the active adapter layer are trained, while the original Transformer layer remains inactive and no parameter update calculations are performed. This improves the computing efficiency of the end-side devices and reduces the memory overhead introduced by storing intermediate parameters.

[0085] In its specific implementation, the collaborative fine-tuning system adopts a hierarchical model segmentation strategy, with the Transformer module as the basic segmentation granularity and deployed on a group of end-side devices according to rules. At the same time, it maintains parameter copies of the input embedding layer (Embedding Layer, Emb) and the output head layer (Head Layer, Hed) on each device, thereby constructing a ring topology with any data source device as the starting and ending nodes, and switching the data source node in a polling manner in different training rounds to achieve effective utilization of the local data of all participating devices.

[0086] like Figure 2 As shown, Figure 2 The figure shows a pipeline parallel training diagram, including end-side devices u1, u2, u3, and u4. During the startup phase, data is forward propagated. Fi represents the forward calculation for data batch i, and Bi represents the backward calculation for data batch i. Before the backward propagation (B1) for the first data batch is completed, the forward propagation for the first, second, third, and fourth data batches has already begun, all based on the same version of the model weights as the first data batch. This may result in different model update directions for the first four data batches, potentially extending convergence time. Furthermore, to unify the model update directions for the four data batches, each device is required to store four versions of the model parameters for average updates. This introduces significant storage overhead, exacerbating resource challenges for end-side devices, which already have limited memory and storage capacity, and can easily lead to fine-tuning failures.

[0087] Step S20 , updating the adjustable adapter layer by layer according to the training status of the large model to be adjusted, to obtain a large model embedded with the updated configuration of the adjustable adapter.

[0088] It should be noted that in this implementation, the adjustable adapters in the task running phase are dynamically configured, the model training status and device resource changes are perceived online, the underlying adapters are activated layer by layer as needed, and the adjustable parameters of the high-importance adapters in the activated state are adaptively increased, so as to save the device resource overhead in the initial training stage without sacrificing the convergence speed and fine-tuning performance.

[0089] In an implementable embodiment, step S20 can comprise: obtaining a connection order of the coding-decoding layers in the large model to be adjusted in the large model; activating the corresponding adjustable adapters layer by layer according to the training state of the large model to be adjusted and the connection order, to obtain activated adjustable adapters; adjusting the width configuration of the activated adjustable adapters, to obtain updated adjustable adapters; and embedding the updated adjustable adapters into the large model to be adjusted, to obtain a large model embedding the updated adjustable adapters.

[0090] It should be noted that the level of the adjustable adapter embedded behind each coding-decoding layer can be determined according to the connection order of the coding-decoding layer in the large model to be adjusted. In view of the progressive learning characteristics of the “from easy to difficult” in the fine-tuning training process, a strategy of dynamically expanding the activated adjustable parameter quantity is adopted, i.e., taking a certain number of training rounds as a period, by judging whether a temporary stagnation point is encountered in the training process (i.e., further training under the current adapter configuration will no longer bring significant accuracy improvement), the adapter parameters of a lower layer are gradually unfrozen from top to bottom, and the bottleneck dimension of part of the important activated adapters is increased. In the present embodiment, a parameter solidification strategy from bottom to top can also be used to shorten the back propagation path, which is not specifically limited in the present embodiment, and the present embodiment is described by taking the parameter unfreezing strategy from top to bottom as an example.

[0091] It can be understood that the depth expansion of the adjustable adapters in each layer, i.e., the activation of the adjustable adapters in each layer, means that at the beginning of the fine-tuning training, the adapter layer parameters of the topmost layer (i.e., the layer closest to the model output) are first unfrozen, and every k training rounds, the adapter parameters of the next layer are unfrozen to participate in the update calculation. Once any parameter position of any adapter layer is unfrozen, it will remain in the activated state until the training converges.

[0092] In an implementable embodiment, the adjusting the width configuration of the activated adjustable adapters to obtain updated adjustable adapters comprises: obtaining gradient information of the current adjustable parameters of the activated adjustable adapters; determining the trace of the Fisher information matrix based on the gradient information of the current adjustable parameters; sorting the activated adjustable adapters based on the trace of the Fisher information matrix to obtain target adapters; and updating the bottleneck dimension of the target adapters based on a preset bottleneck dimension growth step, to obtain updated adjustable adapters.

[0093] It should be noted that in the fine-tuning training process, the Fisher information matrix is a key tool for measuring the importance of parameters, which is based on the gradient information of the model parameters, and can provide a quantitative indicator of the degree of influence of the parameters on the model output. By calculating the trace of the Fisher information matrix (i.e. the sum of the diagonal elements of the matrix), the importance of the adjustable adapter can be determined.

[0094] It can be understood that the gradient information of all parameters in the adjustable adapter layer of the current activation state is calculated, which reflects the contribution of each parameter to the improvement of model performance at the current training stage. Based on these gradient information, the Fisher information matrix can be constructed, and its trace can be calculated. The size of the trace can be used as a measure of the importance of the adapter. The adapter with a larger trace indicates that it has a more significant impact on the model performance. According to the trace of the Fisher information matrix, the adjustable adapters of the activation state are sorted to determine the target adapter, which is the important adapter at the current training stage. Based on the preset bottleneck dimension growth step, the bottleneck dimension of the target adapter is updated. The bottleneck dimension refers to the limitation or adjustment of the number of parameters in the adjustable adapter layer. By gradually increasing the bottleneck dimension, fine-tuning of important parameters can be achieved, thereby further improving the performance of the model without sacrificing convergence speed and fine-tuning performance. The trace of the Fisher information matrix is used to guide the parameter optimization of the adjustable adapter layer, which can effectively balance the efficiency and performance of model training, and ensure efficient and accurate model fine-tuning on the end-side device.

[0095] It should be noted that in the embodiment, the trace of the Fisher information matrix (denoted as tr(F)) can be used to evaluate the training state and guide the configuration decision of the adapter. Alternatively, an approximation of the Fisher information or other measurement indicators can be used to evaluate the training state and the real-time importance of different model parameters or modules for training performance. The present embodiment does not make specific limitations, and the trace of the Fisher information matrix is used as an example for illustration.

[0096] It should be noted that the width expansion of the adapter refers to calculating the trace of the Fisher information matrix (i.e. the tr(F) value) of the currently activated adapter layer based on the gradient information of its current adjustable parameters, and sorting all activated adapters in descending order according to their tr(F) values. The adapters ranked in the top α% are considered important adapters for the current training stage. Under the memory constraints of the end-side device, the bottleneck dimension of the top α% important adapters is linearly increased, and the new adjustable parameters are initialized with a zero-mean Gaussian distribution, as follows:

[0097]

[0098] where γ is the bottleneck dimension growth step, is the maximum allowed bottleneck dimension of the jth adapter under the memory constraint of the end-side device u.

[0099] It is worth noting that the configuration update interval is calculated by calculating the average tr(F) of all adjustable parameters and the average duration of a single training round. The ratio of device calculation delay and point-to-point communication delay is used to evaluate the performance of model convergence and system efficiency under the current adapter configuration strategy, as shown in the following formula:

[0100]

[0101] Among them, TE is the training efficiency value, k is the configuration update interval, is the average duration of a single training round, and r is the number of training rounds.

[0102] When the training efficiency of the current k rounds decreases compared to the previous 2k to k+1 rounds, that is, TE(r-2k:rk)>TE(rk:r), the configuration update interval k is linearly increased to k+Δ to ensure that the system continues to use the current adapter depth and width configuration, avoiding an aggressive increase in the number of adjustable parameters.

[0103] like Figure 3 As shown, Figure 3 The online configuration decision flow chart for the adapter is as follows: when the training round r=1, determine whether r%k=0 is satisfied. If not, perform the rth round of training and r=r+1; if satisfied, evaluate the training efficiency TE(rk:r), sort the importance of the adapter layers, increase the width of the first α% important adapter layers, increase the adapter activation depth, initialize the newly added model parameters, and determine whether TE(r-2k:rk)>TE(rk:r). If so, configure the update interval k to k+Δ. If not, maintain the configuration update interval k.

[0104] like Figure 4 As shown, Figure 4The structure diagram of the Transformer large model structure fused with the serial adapter is shown in the figure. The large model is composed of an input embedding layer (Emb), an output header layer (Hed), and a plurality of Transformer layers. Forward propagation (Forward) and backward propagation (Backward) are performed between the Transformer layers. Each Transformer layer is composed of three residual connections and layer normalization (Add&Norm), an adapter (Adapter), a feed-forward layer (Feed-forward), and a multi-head attention mechanism (Multi-head Attention). Each adapter (Adapter) module is mainly composed of two feed-forward (Feedforward) sublayers. The first feed-forward sublayer (down-project) takes the output of the Transformer block as input, projects the original input dimension d (high-dimensional feature) to m (low-dimensional feature), limits the parameter quantity of the Adapter module by controlling the size of m, and then passes through a nonlinear layer (Nonlinearity). In the output stage, the input dimension is restored by the second feed-forward sublayer (up-project), and m (low-dimensional feature) is mapped back to d (original high-dimensional feature) as the output of the Adapter module.

[0105] In a specific implementation, in the fine-tuning training process, a top-down adapter layer-by-layer unfreezing strategy is adopted. Only the adapter parameters in the active state participate in the update calculation of the backward propagation. The central controller makes online decisions on the adapter configuration strategy, including configuring the top-down unfreezing depth of the adapter and the structure width of each active state adapter, according to the real-time observation of the computing power, memory, and link state of each end-side device in the training process, and the evaluation of the model convergence state. Based on the configuration, for any training iteration, the forward propagation calculation starts from the end-side device holding the data sample processed in the current iteration, generates a feature representation of the data through the Emb layer of the device, and then sequentially passes through each collaborative end-side device in the order of Transformer layer deployment, and finally transmits to the original input device end-side device to generate a model output result. The loss value is calculated based on the corresponding data label. The propagation process forms a ring topology structure. Adjacent end-side devices can transmit the required inter-layer parameters through point-to-point communication methods such as direct communication between devices. Similarly, the backward propagation calculation of this training iteration is performed in reverse along the end-side devices in the ring topology. The inter-layer gradient values are transmitted between adjacent devices in the topology. The backward propagation process stops at the deepest adapter layer under the current unfreezing depth configuration.

[0106] Step S30, performing forward propagation calculation and backward propagation calculation on the large model with the updated configuration of the embedding update adjustable adapter to obtain the large model after parameter adjustment.

[0107] It should be noted that forward propagation and backward propagation are two key steps that are alternately performed during the training process. Forward propagation is responsible for processing input data through the various layers of the model to ultimately produce a prediction result. This process involves the collaborative work of the input embedding layer, the Transformer layer, and the output head layer. In each Transformer layer, the data first undergoes residual connection and layer normalization processing, then dimensionality reduction and dimensionality increase operations are performed through the adapter module, and finally further processing is performed through the feedforward layer and multi-head attention mechanism until the final feature representation is output. Backward propagation, on the other hand, is based on the error calculated by the loss function, and the error signal is propagated backward to update the parameters in the model. This process also needs to go through the various sub-modules of the Transformer layer, including the adapter module, as well as residual connection and layer normalization. Through the gradient descent algorithm, the model parameters are adjusted to reduce the difference between the prediction result and the true value.

[0108] In a feasible implementation, step S20 can include: performing forward propagation calculation and backward propagation calculation on the sub-models of each end-side device based on the task data samples to realize pipeline parallel training and obtain a trained large model; evaluating training convergence according to a loss value of the trained large model to obtain an evaluation result; when the evaluation result is training convergence, taking the trained large model as a parameter-adjusted large model; when the evaluation result is training non-convergence, performing round-robin training on the trained large model until training convergence is achieved to obtain the parameter-adjusted large model.

[0109] It should be noted that in pipeline parallel training, each end-side device is responsible for a different part of the model, allowing the entire training process to be performed in parallel. After receiving the data samples, each end-side device first performs forward propagation calculation, processing the data through the model part it is responsible for. After processing is complete, the data is passed to the next end-side device, and the process continues until all end-side devices have completed their respective model part calculations. After completing its task, each end-side device passes the result to the next device, forming a pipeline job.

[0110] It can be understood that in pipeline parallel training, since each end-side device is executed sequentially, it is necessary to ensure the efficiency and accuracy of data transmission between devices. To achieve this, high-speed network connections are typically used between end-side devices to reduce data transmission time. In addition, to ensure the stability of the training, each end-side device needs to be synchronized to ensure that each device has completed the corresponding calculation task when processing data.

[0111] It is worth mentioning that in the implementation process of pipeline parallel training, the central controller is responsible for coordinating the work of each end-side device. It allocates tasks reasonably according to the computing power, memory and link state of each device, and ensures that each device can work efficiently during the training process. In addition, the central controller will dynamically adjust the strategy of pipeline parallel training according to the evaluation of model convergence state, such as adjusting task allocation, update interval, etc., to improve training efficiency and model performance. Through pipeline parallel training, the training speed of large-scale models can be significantly improved, and the time from training to deployment can be shortened. At the same time, since each end-side device can process data in parallel, distributed computing resources can be fully utilized to improve computing efficiency.

[0112] It is worth mentioning that during the training process, the calculation of loss value reflects the difference between the model prediction result and the true label. When the loss value reaches a lower level, it indicates that the model has well fitted the training data, and thus it can be considered that the training process has converged. In order to accurately evaluate the training convergence, the central controller will periodically calculate the loss value of the entire model and compare it with the preset threshold. If the loss value is lower than the threshold, it is considered that the model has converged, at which point the training can be stopped and the current model can be used as the final parameter-adjusted large model. On the contrary, if the loss value is still high, it indicates that the model has not learned the data characteristics sufficiently and needs to continue training. In this case, the central controller will decide whether to perform round-robin training according to the current training state and resource situation, i.e., training is performed in a loop among the various end-side devices until the model converges. Round-robin training helps to ensure that each end-side device can fully participate in the training process, while avoiding the decline in overall training efficiency due to the computational bottleneck of a single device. In this way, the model can achieve the best training effect under limited computing resources.

[0113] It is worth mentioning that in the fine-tuning training process, the top-down adapter layer-by-layer unfreezing strategy ensures the training efficiency and performance of the model. To further optimize this process, a gradient accumulation-based training method can be introduced, which allows the gradient to be accumulated in each training batch instead of immediately updating the model parameters. In this way, the model can observe the trend of gradient changes over a longer period of time, allowing for more accurate parameter adjustments. Specifically, when a training batch begins, the gradient is initialized to zero. As each training sample is forward-propagated and backward-propagated, the gradient is accumulated on the initial value. Only when all samples in a batch have been processed, the model parameters are updated according to the accumulated gradient. This method is particularly suitable for handling small batches of data, effectively reducing the variance of gradient estimation and improving the stability and convergence speed of the model.

[0114] It can be understood that in the present embodiment, fusion is adopted to fine-tune the adapter, an adapter layer is embedded after each Transformer layer, the bottom adapter is kept in an inactive state at the beginning of fine-tuning, and since only the adapter parameters in the unfrozen state are updated by gradient descent during fine-tuning, the characteristics of early stopping of backward propagation calculation are utilized to perform forward propagation calculation on multiple batches of data in advance, pipeline parallel computing among multiple devices is realized, and other parameter efficient fine-tuning techniques such as LoRA, prompt fine-tuning, etc. can also be used, which are not specifically limited in the present embodiment, and the adapter fine-tuning technique is taken as an example for description.

[0115] It can be understood that the on-demand layer-by-layer parameter activation strategy shortens the reverse propagation calculation path, and the characteristics of fixed bottom model parameters are utilized to start forward propagation calculation of a new round of training iteration in advance, which promotes parallel computing among devices without introducing model obsolescence and increasing parameter copy storage overhead, and can realize significant acceleration of the fine-tuning process.

[0116] In a feasible embodiment, the sub-models of the end-side devices are subjected to forward propagation calculation and backward propagation calculation based on the task data samples to realize pipeline parallel training and obtain a trained large model, comprising: determining a forward propagation path and a backward propagation path based on the end-side devices; performing forward propagation training on the sub-models of the end-side devices along the forward propagation path and performing backward propagation training on the sub-models of the end-side devices along the backward propagation path based on the task data samples to realize pipeline parallel training and obtain a trained large model.

[0117] It should be noted that the forward propagation path and the backward propagation path define the order of data flowing among the end-side devices. The forward propagation path starts from the input data, and sequentially passes through the sub-models of each end-side device until the last device completes the forward propagation calculation. The backward propagation path starts from the last end-side device, and passes through each sub-model in reverse until the first device completes the backward propagation calculation. Such path setting ensures efficient data transmission among the end-side devices, while ensuring the synchronization of the training process.

[0118] It can be understood that for any training iteration, the sampling device of the sample data is referred to as an iteration "initiating node", and the input embedding layer model of the device inputs the batch data of the current iteration; the end-side device with the lowest unfrozen adapter layer is referred to as an iteration "aborting node", and the backward propagation calculation of the current iteration stops at the activated adapter layer carried by the device.

[0119] In a specific implementation, in the initialization phase, each device uploads its state information to the central controller, including the communication rates of all available point-to-point links, the self-calculated speed and memory constraints of each device. The central controller determines the model partitioning and allocation strategy according to the collected system state information, and distributes the pre-trained layer parameters to the corresponding end-side devices. The central controller selects and notifies the first "initiating node" in each training round. At the same time, the central controller also sends the necessary training settings (including learning rate, local iteration number and batch size) to the end-side devices to initialize the training. In each training iteration, the selected "initiating node" samples a small batch of data from its local data set, inputs it into its local input embedding layer, and sends the output intermediate parameters to the corresponding next end-side device according to the arrangement order of the Transformer block. The intermediate results generated by each end-side device sub-model are passed along the ring topology, and the output of the last Transformer block is sent back to the "initiating node" device to calculate the training loss, thereby completing the forward propagation of the data batch. Then, the "initiating node" starts the backpropagation calculation based on the locally calculated gradient, and the "initiating node" updates its output head layer parameters, and the gradient is passed between end-side devices along the reverse ring topology, and the end-side devices carrying the activation adapter layer in the reverse propagation path update their adapter layer parameters accordingly, and the backpropagation calculation stops at the "abort node".

[0120] It is worth noting that in the present embodiment, since the underlying model does not contain activation parameters that need to be updated by backpropagation, the training process allows the "initiating node" device to continuously sample small batches of data and input them into the distributedly deployed large model before the end of the forward and backward propagation calculations of the current iteration, to start the forward propagation of a new training iteration, that is, the distributed training system can perform multiple iterations of forward / backward propagation calculations at the same time, and different devices may simultaneously undertake the calculation tasks of different training iterations, achieving a pipeline parallel training process between end-side devices with batch data as the granularity.

[0121] It is worth noting that after completing a predetermined number of iterations, the "initiating node" device reports the local loss value to the central controller to evaluate the training convergence, while selecting the next "initiating node" with the best channel quality with itself, and passing the latest parameters of the local output head layer to update its head layer. When all end-side devices in the system have played the role of "initiating node", the current training round ends, and it is determined whether it is an adapter configuration update round. If not, a new round of training is entered; if so, the central controller obtains the current Fisher matrix and other training state information of the sub-model from each end-side device, as well as the real-time communication, computing power, memory and other system state information of the device, determines the adapter update configuration, and sends the configuration decision to each end-side device. The end-side device initializes the newly activated parameters according to the latest adapter configuration, and repeats the above process until convergence, i.e., the model fine-tuning is completed.

[0122] As shown in Figure 5 , Figure 5 is a schematic diagram of multi-device collaborative fine-tuning training. In the figure, four devices u1, u2, u3, and u4 collaboratively support the fine-tuning task of a large model with 12 Transformer modules. The Transformer layers are loaded onto the u1, u2, u3, and u4 devices in the order of 4:5:2:3, and it is assumed that the adapter unfreezing depth output by the adapter online configuration scheme is 3. It is assumed that the sampling starts from the local sample data of u1. The first training iteration is performed along the path u1→u2→u3→u4→u1 for forward propagation, and along the path u1→u4 for backward propagation for training. The above iteration process is repeated. When the u1-side data training is completed h times, the latest Hed layer parameters of u1 are transmitted to u2 to update the u2-side Hed layer. Subsequently, u2 samples a new batch of data from its local data set and starts u2-side data training. After all end-side devices are traversed once through the above training process, it is considered as a training round. The polling collaborative training process is repeated until the model converges.

[0123] In the above multi-device collaborative fine-tuning training process, two parallel processing operations are further integrated: 1) In different iterations of a device-side data training, the backward propagation update calculation process of the model and the forward propagation calculation process before the adapter unfreezing layer are independent of each other and can be processed in parallel; 2) In the iterations of the first and second device-side data training, the backward propagation update calculation process of the model Transformer layer and the Hed layer parameter transmission process are independent of each other and can be executed in parallel. The above two types of calculation-computation and calculation-transmission parallel processing help to improve the utilization of device resources and accelerate the training process.

[0124] The embodiment provides a large model adjusting method, and the embodiment can effectively improve the large model parameter adjusting efficiency and save device resources by first performing segmented deployment on a large model to be adjusted to obtain sub-models deployed on different end-side devices, wherein each sub-model comprises different numbers of codec layers and adjustable adapters, and each codec layer is embedded with a corresponding adjustable adapter; updating the adjustable adapters layer by layer according to the training state of the large model to be adjusted to obtain a large model embedded with the updated adjustable adapters; and performing forward propagation calculation and backward propagation calculation on the large model embedded with the updated adjustable adapters to obtain a large model after parameter adjustment.

[0125] In conclusion, the embodiment can accurately adapt to corresponding end-side devices by performing segmented deployment on a large model to be adjusted, obtain sub-models deployed on different end-side devices, update the adjustable adapters in each sub-model layer by layer, effectively shorten the reverse propagation path, and then perform forward propagation and backward propagation, so that the large model is fine-tuned through multiple devices, and the adjusted large model is quickly and efficiently obtained, thereby overcoming the technical defects that the model cannot be fine-tuned through multiple devices and the model fine-tuning fails due to limited device resources, and effectively improving the large model parameter adjusting efficiency and saving device resources.

[0126] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as the above embodiment one can refer to the above introduction, and the subsequent will not be repeated. On this basis, please refer to Figure 6 , the step S10 further comprises steps S101-S104:

[0127] Step S101, determining the number of codec layers allocated to each end-side device according to the memory capacity of each end-side device.

[0128] It should be noted that the memory capacity of the end-side device is one of the important factors affecting its processing capability, and in the embodiment, by considering the memory capacity of the end-side device, the codec layers can be more reasonably allocated to ensure that each device can efficiently perform its assigned tasks, and avoid performance bottlenecks or training failures caused by insufficient memory.

[0129] It can be understood that in the model segmentation and allocation process, the minimum training delay is taken as the target, the differentiated computing speed, memory capacity and link rate among the cooperative end-side devices are considered, the device ring topology construction decision is made based on the graph representation and optimization technology, and the allocation and deployment scheme of the Transformer module in each device is determined to shorten the idle waiting time of the heterogeneous devices in the fine-tuning training process and improve the convergence acceleration advantage brought by the pipeline parallel training.

[0130] In a specific implementation, before starting the fine-tuning task, the center controller first determines the large model partition deployment strategy according to the available computing power, memory capacity, point-to-point communication rate and other information of the end-side devices in the system, including which Transformer layers the model is partitioned to form multiple groups of sub-models containing consecutive Transformer layers, and which device each sub-model is allocated to for processing. The center controller broadcasts the partition deployment strategy and distributes each sub-model parameter to the corresponding end-side device according to the strategy to perform the corresponding forward and backward propagation calculations. In addition, each Transformer layer is connected in series with an adapter module with a specific initialization structure and parameters, which is delivered to the corresponding end-side device by the center controller together with the associated Transformer layer parameters. In addition to the allocated sub-models, each end-side device also locally deploys a copy of Emb and Hed to perform feature representation, generate model output results, loss value calculation and output head layer parameter update calculation on its local sample data.

[0131] In a feasible implementation, step S101 can include: obtaining a plurality of end-side devices and corresponding forward propagation calculation delays, and sorting the end-side devices based on the forward propagation calculation delays to obtain sorted end-side devices; determining a number of encoding and decoding layers threshold for each end-side device according to the memory capacity of each end-side device; and determining the number of encoding and decoding layer allocations for each end-side device according to the number of encoding and decoding layer threshold and the sorted end-side devices.

[0132] It should be noted that for any collaborative fine-tuning task, model partition deployment is performed once before system construction, and the end-side devices and their direct transmission links are modeled as a directed graph G=(U,A), where U is the device vertex set, and A is the directed edge set, and its element a=(u,u') represents the directed transmission link from device u to device u'. The link transmission delay is defined as and the forward propagation calculation delay of a Transformer module executed on a device is τ tr (u,l), the forward propagation calculation delay of the end-side device is a manifestation of the computing power of the end-side device. In this embodiment, by considering the forward propagation calculation delay of the end-side device, the encoding and decoding layers can be more reasonably allocated to ensure that each device can efficiently perform its assigned task and avoid performance bottlenecks or training failures due to excessively long calculation delays.

[0133] It can be understood that according to the available memory capacity of each device and the size of each Transformer block, the number of Transformer blocks allocated to each device is optimized to minimize the total calculation time of all devices.

[0134] In a specific implementation, the end-side devices are sorted in descending order according to their computing capabilities (i.e., forward propagation calculation delay); for each sorted end-side device, the maximum number of assignable Transformer blocks, i.e., the threshold of the number of encoding and decoding layers, is calculated based on its memory capacity and the memory requirement of a single Transformer block; the number of unassigned Transformer blocks is updated, the number of Transformer blocks assigned to the next device is determined according to the sorting, and the process is repeated until all devices are traversed; if there are still unassigned Transformer blocks after the traversal ends, it indicates that the edge devices participating in the cooperative fine-tuning cannot carry the target large model, and the model partitioning and deployment algorithm terminates.

[0135] It is worth noting that the Transformer block number allocation optimization process allows computing power or memory-limited partial terminals to be allocated without Transformer blocks, only carrying input embedding layers (Emb) and output header layers (Hed), responsible for processing input data and calculating loss, and not participating in forward and backward propagation calculations of Transformer blocks.

[0136] Step S102: partitioning the large model to be adjusted into multiple sub-models according to the number of encoding and decoding layers.

[0137] It should be noted that according to the number of Transformer blocks corresponding to each end-side device, the size and complexity of each sub-model can be dynamically adjusted to adapt to the computing power of different devices. For example, for edge devices with strong computing power, more Transformer blocks can be allocated to make them undertake more complex tasks; while for devices with weak computing power, fewer Transformer blocks are allocated, mainly responsible for basic input and output processing.

[0138] It can be understood that if the large model to be adjusted in the embodiment has 12 layers of Transformer modules, and the corresponding number of Transformer blocks of the edge-side devices is 4, 5, 2, and 3 respectively, the 12 layers of Transformer modules can be divided into four sub-models, i.e., a sub-model containing 4 Transformer blocks, a sub-model containing 5 Transformer blocks, a sub-model containing 2 Transformer blocks, and a sub-model containing 3 Transformer blocks. In this way, each edge-side device can effectively participate in the processing of the large model according to its own computing power, while ensuring the efficiency and accuracy of the overall task. In addition, this dynamic allocation mechanism can also effectively improve the flexibility and scalability of the system. In actual application, as the performance of the device improves or the task demand changes, the system can flexibly adjust the number of Transformer blocks of each sub-model to adapt to new demands. For example, when the computing power of an edge device is enhanced, more Transformer blocks can be allocated to the device to enable it to handle more complex tasks, thereby fully utilizing the computing resources of the device.

[0139] In step S103, a target Hamilton circuit is obtained.

[0140] It should be noted that a Hamilton circuit is a special path in graph theory that passes through each vertex in the graph exactly once and eventually returns to the starting point. In the present embodiment, the purpose of obtaining the target Hamilton circuit is to determine the data flow order between sub-models, to ensure that the data dependency relationship between sub-models is correctly handled when the large model is segmented, and to ensure the running efficiency and accuracy of the entire model.

[0141] In a feasible embodiment, step S103 can include: obtaining transmission links between each edge-side device, and constructing a directed graph based on each edge-side device and the transmission links, wherein the directed graph includes a plurality of original nodes; adding virtual nodes in the directed graph, and constructing directed edges based on the virtual nodes and each original node; obtaining the weights of each directed edge, and generating a directed weighted graph based on the weights of each directed edge; determining the target Hamilton circuit based on the directed weighted graph.

[0142] It should be noted that the edge-side devices and the direct transmission links therebetween are modeled as a directed graph G=(U,A), where U is a set of device nodes, and A is a set of directed edges, and an element a=(u,u') of A represents a directed transmission link from device u to device u', i.e., the original nodes in the directed graph are the edge-side devices, and the directed edges are the directed transmission links between devices u and u'.

[0143] It can be understood that it is necessary to check whether there are devices that have not been assigned Transformer blocks. If so, the nodes corresponding to these devices and their associated edges are removed from the directed graph G; a virtual node u is added to the directed graph G. end , and the virtual node u end Establish directed edges with all nodes in the G directed graph; calculate the weights of each directed edge to form a new graph, namely the directed weighted graph G′(U - ,A - ,W), the weight of each directed edge is as follows:

[0144]

[0145] Among them, Weight(u,u') is the weight of the directed edge, U － is a node set of a directed weighted graph, u and u' are device nodes on the end side, u end For virtual nodes.

[0146] In a feasible implementation, determining the target Hamiltonian circuit based on the directed weighted graph includes: taking a virtual node in the directed weighted graph as a starting point and performing path planning to generate a path table; determining the path length from each visited node in the visited node set to an unvisited node based on the path table; determining whether a reference Hamiltonian circuit exists based on the path length; and when the reference Hamiltonian circuit exists, removing the virtual node and reconstructing the path through a backtracking table to generate a target Hamiltonian circuit.

[0147] It should be noted that the dynamic programming method is used to find the shortest directed Hamiltonian circuit based on the directed weighted graph. The circuit passes through each vertex once along the directed edge and finally returns to the starting vertex, thereby constructing a The path table Dis(V,v), where V is the set of visited vertices, v is the current vertex, and each element in the table represents the path from the virtual vertex u end Starting from, passing through all vertices of set V, and finally reaching vertex v, the shortest path length (i.e., the sum of edge weights)

[0148] It is understandable that for each subset and vertex v∈V, calculate the distance from v to the unvisited vertex The path length is calculated and the path table and backtracking table are updated to obtain Dis(V′,v′) and Back(V′,v′). After calculating the shortest paths of all subsets, find the path that can return to the virtual vertex u. end The shortest Hamiltonian circuit; if the optimal path is found, the complete path is reconstructed through the backtracking table Back(V′,v′) and the virtual vertex u is removed end To form the final path, that is, the target Hamiltonian circuit.

[0149] It is worth noting that the backtracking table is an important data structure for recording the predecessor nodes of each node in the path planning process. Through the backtracking table, the complete Hamiltonian circuit can be reconstructed by tracing back from the end point to the starting point. In the process of dynamic programming, the construction of the backtracking table is completed step by step. Whenever the shortest path from a virtual vertex to a certain visited vertex is determined, the predecessor node of the vertex in the backtracking table is updated.

[0150] Step S104, based on the target Hamiltonian circuit, sequentially deploying a plurality of sub-models on the corresponding end-side devices to obtain sub-models deployed on different end-side devices.

[0151] It should be noted that the target Hamiltonian circuit is a path that covers all end-side device nodes and each node is visited only once, and the shortest path returns to the starting point. After determining the target Hamiltonian circuit, the sub-models can be deployed to the corresponding end-side devices in the order of the circuit.

[0152] In a specific implementation, first, the deployment order of the sub-models is determined, that is, arranged according to the node order of the target Hamiltonian circuit. Then, for each end-side device, according to its position in the target Hamiltonian circuit, the corresponding sub-model is assigned. In the assignment process, factors such as the computing power, storage space and network bandwidth of the device need to be considered to ensure that the sub-model can run normally on the device and meet the performance requirements.

[0153] In a feasible implementation, step S104 can include: determining the end-side device arrangement order based on the target Hamiltonian circuit; obtaining the model parameters of each sub-model; and sequentially sending the model parameters to the corresponding end-side devices according to the end-side device arrangement order to obtain sub-models deployed on different end-side devices.

[0154] It should be noted that the end-side device arrangement order determined according to the target Hamiltonian circuit can ensure the orderliness and efficiency of the sub-model deployment. The model parameters of each sub-model are obtained, which include but are not limited to the model's weight, bias, activation function and other information. These parameters are the basis for the sub-model to perform inference and learning.

[0155] It can be understood that the model parameters are sequentially sent to the corresponding end-side devices according to the arrangement order of the end-side devices. In the sending process, compression techniques can be used to reduce data transmission volume while ensuring data integrity and accuracy.

[0156] As shown in FIG. 4, the target Hamiltonian circuit is determined based on the path planning result of the first path planning process. Figure 7 Figure 7 ​A model segmentation deployment decision flowchart is constructed, a device connection topology graph is constructed, the maximum number Trm is allocated in descending order of device computing power, it is judged whether there is an unallocated Trm, if there is, the device connection graph is updated, if not, the edge weight of the connection graph is determined, the shortest Hamiltonian circuit is determined, and then the Trm deployment strategy is determined.

[0157] In this embodiment, by accurately determining the number of codec layers allocated according to the memory capacity of each end-side device and further model segmentation, and then deploying the sub-models in the corresponding end-side devices according to the target Hamiltonian circuit, the idle waiting time of the heterogeneous devices in the fine-tuning training process can be effectively shortened, and the convergence speed of the training is improved.

[0158] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the large model adjustment method of the present application. More simple transformations based on this technical concept are within the scope of protection of the present application.

[0159] The present application also provides a large model adjustment device, please refer to Figure 8 , the large model adjustment device comprises:

[0160] The deployment module 10 is configured to segment and deploy the large model to be adjusted to obtain sub-models deployed on different end-side devices, wherein each sub-model comprises a different number of codec layers and adjustable adapters, and each codec layer is embedded with a corresponding adjustable adapter.

[0161] The update module 20 is configured to update the adjustable adapters layer by layer according to the training state of the large model to be adjusted to obtain a large model embedded with the updated configuration of the adjustable adapters.

[0162] The propagation module 30 is configured to perform forward propagation calculation and backward propagation calculation on the large model embedded with the updated configuration of the adjustable adapters to obtain a large model after parameter adjustment.

[0163] The present application provides a large model adjustment device, which segments and deploys a large model to be adjusted to obtain sub-models deployed on different end-side devices, wherein each sub-model comprises a different number of codec layers and adjustable adapters, and each codec layer is embedded with a corresponding adjustable adapter; the adjustable adapters are updated layer by layer according to the training state of the large model to be adjusted to obtain a large model embedded with the updated configuration of the adjustable adapters; and the forward propagation calculation and the backward propagation calculation are performed on the large model embedded with the updated configuration of the adjustable adapters to obtain a large model after parameter adjustment, which can effectively improve the large model parameter adjustment efficiency and save device resources.

[0164] To sum up, the embodiment can accurately adapt to the corresponding end-side device by splitting and deploying the to-be-adjusted large model, obtain sub-models deployed on different end-side devices, update the adjustable adapters in each sub-model layer by layer, effectively shorten the back propagation path, and thus perform forward propagation and backward propagation, so as to fine-tune the large model through multiple devices and quickly and efficiently obtain the adjusted large model. The technical defects that the model fine-tuning cannot be performed through multiple devices and the model fine-tuning fails due to limited device resources are overcome, the large model parameter adjustment efficiency can be effectively improved, and device resources can be saved.

[0165] Optionally, the deployment module 10 is further configured to determine the number of codec layers of each end-side device according to the memory capacity of each end-side device; split the to-be-adjusted large model into multiple sub-models according to the number of codec layers; obtain a target Hamilton circuit; and deploy the multiple sub-models on the corresponding end-side devices in sequence based on the target Hamilton circuit to obtain sub-models deployed on different end-side devices.

[0166] Optionally, the deployment module 10 is further configured to obtain multiple end-side devices and corresponding forward propagation calculation delays, sort the end-side devices based on the forward propagation calculation delays to obtain sorted end-side devices, determine the number of codec layers of each end-side device according to the memory capacity of each end-side device, and determine the number of codec layers of each end-side device according to the number of codec layers threshold and the sorted end-side devices.

[0167] Optionally, the deployment module 10 is further configured to obtain a transmission link between each end-side device, and construct a directed graph according to each end-side device and the transmission link, wherein the directed graph includes multiple original nodes; add a virtual node in the directed graph, and construct a directed edge based on the virtual node and each original node; obtain the weight of each directed edge, and generate a directed weighted graph based on the weight of each directed edge; and determine a target Hamilton circuit based on the directed weighted graph.

[0168] Optionally, the deployment module 10 is further configured to take the virtual node in the directed weighted graph as a starting point and perform path planning to generate a path table, determine the path length from each visited node in a set of visited nodes to an unvisited node based on the path table, determine whether a reference Hamilton circuit exists according to the path length, remove the virtual node and reconstruct a path through a backtracking table to generate a target Hamilton circuit when the reference Hamilton circuit exists.

[0169] Optionally, the deployment module 10 is further configured to determine an end-side device arrangement order based on the target Hamilton circuit, obtain model parameters of each sub-model, and sequentially send the model parameters to the corresponding end-side device according to the end-side device arrangement order to obtain sub-models deployed on different end-side devices.

[0170] Optionally, the updating module 20 is further configured to obtain a connection order of the codec layers in the to-be-adjusted large model in each sub-model; activate the corresponding adjustable adapter layer by layer according to the training state of the to-be-adjusted large model and the connection order to obtain an activated adjustable adapter; adjust the width configuration of the activated adjustable adapter to obtain an updated configurable adjustable adapter; and embed the updated configurable adjustable adapter into the to-be-adjusted large model to obtain a large model embedded with the updated configurable adjustable adapter.

[0171] Optionally, the updating module 20 is further configured to obtain gradient information of a current adjustable parameter of the activated adjustable adapter; determine a trace of a Fisher information matrix based on the gradient information of the current adjustable parameter; sort the activated adjustable adapter based on the trace of the Fisher information matrix to obtain a target adapter; and perform bottleneck dimension updating on the target adapter based on a preset bottleneck dimension growth step to obtain the updated configurable adjustable adapter.

[0172] Optionally, the propagation module 30 is further configured to perform forward propagation calculation and backward propagation calculation on the sub-models of each end-side device based on task data samples to implement pipeline parallel training to obtain a trained large model; evaluate training convergence based on a loss value of the trained large model to obtain an evaluation result; when the evaluation result is training convergence, take the trained large model as a parameter-adjusted large model; and when the evaluation result is training non-convergence, perform polling training on the trained large model until training convergence is achieved to obtain the parameter-adjusted large model.

[0173] Optionally, the propagation module 30 is further configured to determine a forward propagation path and a backward propagation path based on the end-side devices; perform forward propagation training on the sub-models of each end-side device along the forward propagation path and perform backward propagation training on the sub-models of each end-side device along the backward propagation path based on task data samples to implement pipeline parallel training to obtain a trained large model.

[0174] The large model adjustment device provided in the present application adopts the large model adjustment method in the above embodiments, and can solve the technical problem of how to effectively improve the large model parameter adjustment efficiency and save device resources. Compared with the prior art, the large model adjustment device provided in the present application has the same beneficial effects as the large model adjustment method provided in the above embodiments, and other technical features in the large model adjustment device are the same as the features disclosed in the above embodiments, which will not be repeated here.

[0175] The application provides a large model adjustment device, comprising: at least one processor; and a memory connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the large model adjustment method in the above-mentioned embodiment I.

[0176] Reference will now be made to the following description Figure 9 which shows a structural schematic diagram of a large model adjustment device suitable for implementing embodiments of the application. The large model adjustment device in the embodiments of the application can include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 9 The large model adjustment device shown is merely an example and should not impose any limitation on the functions and use range of the embodiments of the application.

[0177] As shown in Figure 9 , the large model adjustment device can include a processing apparatus 1001 (such as a central processor, a graphics processor, or the like) that can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 1002 or loaded from a storage apparatus 1003 into a random access memory (RAM) 1004. In the RAM 1004, various programs and data required for operation of the large model adjustment device are also stored. The processing apparatus 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: input apparatuses 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, and the like; output apparatuses 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, and the like; the storage apparatus 1003 including, for example, a magnetic tape, a hard disk, and the like; and a communication apparatus 1009. The communication apparatus 1009 can allow the large model adjustment device to communicate with other devices wirelessly or by wire to exchange data. Although a large model adjustment device having various systems is shown in the figure, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be alternatively implemented or provided.

[0178] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for executing the method shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiments disclosed in the present application are executed.

[0179] The large model adjustment device provided by the present application adopts the large model adjustment method in the above-mentioned embodiments, and can solve the technical problem of how to effectively improve the large model parameter adjustment efficiency and save device resources. Compared with the prior art, the large model adjustment device provided by the present application has the same beneficial effects as the large model adjustment method provided by the above-mentioned embodiments, and other technical features in the large model adjustment device are the same as the features disclosed in the previous embodiment method, which will not be repeated here.

[0180] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the description of the above-mentioned embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0181] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0182] The present application provides a computer readable storage medium having stored thereon computer readable program instructions (i.e. computer program) for executing the large model adjustment method in the above-mentioned embodiments.

[0183] The computer readable storage medium provided in the application may be, for example, a U disk, but is not limited to an electric, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination thereof. More specific examples of the computer readable storage medium may include, but are not limited to, an electric connection with one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the embodiment, the computer readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer readable storage medium can be transmitted by any suitable medium, including but not limited to an electric wire, an optical cable, an RF (Radio Frequency), and the like, or any suitable combination thereof.

[0184] The above computer readable storage medium may be contained in the large model adjustment device, or may exist separately without being assembled into the large model adjustment device.

[0185] The above computer readable storage medium carries one or more programs, which, when executed by the large model adjustment device, cause the large model adjustment device to: split and deploy a large model to be adjusted to obtain sub-models deployed on different end-side devices, wherein each sub-model includes a different number of codec layers and an adjustable adapter, and each codec layer is embedded with a corresponding adjustable adapter; update the adjustable adapters layer by layer according to the training state of the large model to be adjusted to obtain a large model embedded with updated adjustable adapters; and perform forward propagation calculation and backward propagation calculation on the large model embedded with the updated adjustable adapters to obtain a large model after parameter adjustment.

[0186] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0187] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of present application. In this regard, each block in the flow diagrams or block diagrams can represent a module, a procedure, or a part of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may

[0188] The modules involved in the embodiments of the present application can be implemented in the form of software or in the form of hardware. Among them, the name of the module does not constitute a limitation to the unit itself in some cases.

[0189] The readable storage medium provided by the present application is a computer readable storage medium, which stores computer readable program instructions (i.e. computer program) for executing the above-mentioned large model adjustment method, and can solve the technical problem of how to effectively improve the large model parameter adjustment efficiency and save device resources. Compared with the prior art, the computer readable storage medium provided by the present application has the same beneficial effects as the large model adjustment method provided by the above-mentioned embodiments, which will not be repeated here.

[0190] The application further provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the large model adjustment method as described above.

[0191] The computer program product provided by the application can effectively improve the large model parameter adjustment efficiency and save device resources. Compared with the prior art, the beneficial effects of the computer program product provided by the application are the same as those of the large model adjustment method provided by the above-mentioned embodiments, and are not described here.

[0192] The above-mentioned embodiments are only part of the embodiments of the application, and do not limit the patent scope of the application. Any equivalent structural transformation, direct / indirect application in other related technical fields, or direct / indirect application in other related technical fields based on the technical concept of the application and the content of the specification and drawings are included in the patent protection scope of the application.

Claims

1. A large model adjustment method, characterized in that: The method comprises: The large model to be adjusted is split and deployed to obtain sub-models deployed on different end-side devices. Each sub-model includes a different number of codec layers and adjustable adapters, and each codec layer is embedded with a corresponding adjustable adapter. Obtaining the connection order of the codec layers in each sub-model in the large model to be adjusted; updating the adjustable adapter layer by layer according to the training state of the large model to be adjusted and the connection sequence, to obtain a large model in which the adjustable adapter with the updated configuration is embedded; Performing forward propagation calculation and backward propagation calculation on the large model of the adjustable adapter embedded with the updated configuration to obtain the large model with adjusted parameters; The large model to be adjusted is divided and deployed to obtain multiple sub-models and corresponding end-side devices, including: Determine the number of codec layers to be allocated to each device based on the memory capacity of each device. Dividing the large model to be adjusted into multiple sub-models according to the number of codec layer allocations; Get the target Hamiltonian circuit; Based on the target Hamiltonian circuit, the multiple sub-models are sequentially deployed on corresponding end-side devices to obtain sub-models deployed on different end-side devices.

2. The method according to claim 1, wherein The step of determining the number of codec layers allocated to each end-side device according to the memory capacity of each end-side device includes: Acquire multiple end-side devices and corresponding forward propagation calculation delays, and sort the end-side devices based on the forward propagation calculation delays to obtain sorted end-side devices; Determine the threshold for the number of codec layers on each end-side device based on the memory capacity of each end-side device; The number of codec layers allocated to each end-side device is determined according to the codec layer number threshold and the sorted end-side devices.

3. The method according to claim 1, wherein The obtaining of the target Hamiltonian circuit comprises: Acquire transmission links between each end-side device, and construct a directed graph based on each end-side device and the transmission links, wherein the directed graph includes a plurality of original nodes; Adding a virtual node to the directed graph, and constructing directed edges based on the virtual node and each original node; Obtaining the weight of each directed edge, and generating a directed weighted graph based on the weight of each directed edge; A target Hamiltonian circuit is determined based on the directed weighted graph.

4. The method according to claim 3, wherein Determining a target Hamiltonian circuit based on the directed weighted graph includes: Taking a virtual node in the directed weighted graph as a starting point and performing path planning to generate a path table; Determine the path length from each visited node in the visited node set to an unvisited node based on the path table; determining whether a reference Hamiltonian circuit exists according to the path length; When the reference Hamiltonian circuit exists, the virtual nodes are removed and the path is reconstructed through the backtracking table to generate a target Hamiltonian circuit.

5. The method according to claim 1, wherein The step of sequentially deploying the plurality of sub-models on corresponding end-side devices based on the target Hamiltonian circuit to obtain sub-models deployed on different end-side devices includes: Determining an arrangement order of the end-side devices based on the target Hamiltonian circuit; Get the model parameters of each sub-model; The model parameters are sent to the corresponding end-side devices in sequence according to the arrangement order of the end-side devices, so as to obtain sub-models deployed on different end-side devices.

6. The method according to claim 1, wherein The step of updating the adjustable adapter layer by layer according to the training state of the large model to be adjusted and the connection sequence to obtain a large model in which the adjustable adapter with the updated configuration is embedded comprises: activating the corresponding adjustable adapters layer by layer according to the training state of the large model to be adjusted and the connection sequence to obtain an adjustable adapter in an activated state; Adjusting the width configuration of the activated adjustable adapter to obtain an adjustable adapter with an updated configuration; The adjustable adapter after the updated configuration is embedded into the large model to be adjusted to obtain the large model in which the adjustable adapter after the updated configuration is embedded.

7. The method according to claim 6, wherein The adjusting the width configuration of the activated adjustable adapter to obtain an adjustable adapter with an updated configuration includes: Obtaining gradient information of current adjustable parameters of the adjustable adapter in the activated state; determining a trace of a Fisher information matrix based on the gradient information of the current adjustable parameter; sorting the adjustable adapters in the activation state based on the trace of the Fisher information matrix to obtain a target adapter; The bottleneck dimension of the target adapter is updated based on a preset bottleneck dimension growth step to obtain an adjustable adapter with updated configuration.

8. The method according to claim 1, wherein The step of performing forward propagation calculation and backward propagation calculation on the large model of the adjustable adapter embedded with the updated configuration to obtain the large model with adjusted parameters includes: Based on the task data samples, forward propagation and backward propagation calculations are performed on the sub-models of each end-side device to achieve pipeline parallel training and obtain the trained large model; Evaluate the training convergence according to the loss value of the trained large model to obtain an evaluation result; When the evaluation result indicates that the training has converged, the trained large model is used as the large model after parameter adjustment; When the evaluation result indicates that the training has not converged, the trained large model is subjected to round-robin training until the training converges, thereby obtaining a large model with adjusted parameters.

9. The method according to claim 8, wherein The forward propagation calculation and backward propagation calculation are performed on the sub-model of each end-side device based on the task data sample to implement pipeline parallel training to obtain the trained large model, including: Determining a forward propagation path and a backward propagation path based on the end-side device; Based on the task data samples, forward propagation training is performed on the sub-model of each end-side device along the forward propagation path, and backward propagation training is performed on the sub-model of each end-side device along the backward propagation path to realize pipeline parallel training and obtain a trained large model.

10. A large model adjustment device, characterized in that: The large model adjustment device executes the large model adjustment method according to any one of claims 1 to 9, and the large model adjustment device includes: A deployment module is used to split the large model to be adjusted and deploy it, obtaining sub-models deployed on different end-side devices. Each sub-model includes a different number of codec layers and adjustable adapters, and each codec layer is embedded with a corresponding adjustable adapter. An updating module, configured to update the adjustable adapter layer by layer according to the training state of the large model to be adjusted, to obtain a large model in which the adjustable adapter with the updated configuration is embedded; The propagation module is used to perform forward propagation calculation and backward propagation calculation on the large model of the adjustable adapter embedded with the updated configuration to obtain the large model with adjusted parameters.

11. A large model adjustment device, characterized in that: The large model adjustment device includes: a memory, a processor, and a large model adjustment program stored in the memory and executable on the processor, wherein the large model adjustment program is configured to implement the large model adjustment method according to any one of claims 1 to 9.

12. A storage medium, characterized in that: The storage medium stores a large model adjustment program, which, when executed by a processor, implements the large model adjustment method according to any one of claims 1 to 9.

Images