Model deployment method, device and computer program product for implementing a semiconductor response system

Abstract

The application relates to the technical field of artificial intelligence, and in particular provides a model deployment method, device and computer program product for realizing a semiconductor response system. The method comprises the following steps: performing serialization processing on structured data and unstructured data obtained from a local wafer foundry, obtaining serialized data, processing the serialized data to obtain a training set, performing self-supervised training on a local model by using the training set and with the aid of a pre-training model file, obtaining a retrieval large language model, and deploying the retrieval large language model to a local data center server cluster of the wafer foundry. In the production and manufacturing process at the local end, manufacturing engineers at the production site can seek solutions from the semiconductor response system deployed in the local data center, compared with the non-artificial intelligence technology of a traditional wafer foundry, the time for waiting for software engineers, equipment engineers, process engineers and vendor engineers to arrive at the site is reduced, and a large amount of manpower and material resources is not consumed.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus and computer program product for deploying a model of a semiconductor response system. Background Technology

[0002] The semiconductor wafer manufacturing process involves a large number of steps, and the parameters related to these steps are usually pre-configured by software engineers into the production equipment. The equipment then processes the wafers according to these parameters.

[0003] In related technologies, it is common for the process results of machine equipment to fail to meet requirements. The current common solution is for manufacturing engineers on the production floor to communicate the problem and requirements to software engineers, who then go to the site to try to solve it on-site. This method takes anywhere from minutes to several hours and consumes a lot of manpower, thus delaying the wafer manufacturing process. Summary of the Invention

[0004] This application is made in view of the above-mentioned problems. This application provides a method, apparatus, and computer program product for deploying a model of a semiconductor response system.

[0005] According to the first aspect of this application, a method for deploying a model of a semiconductor response system is provided, comprising:

[0006] Obtain large datasets for training a general-purpose large language model based on a self-attention mechanism architecture;

[0007] Train the general large language model and obtain the pre-trained model file;

[0008] Obtain structured and unstructured data locally from the wafer foundry for training and deploying local models;

[0009] The structured data and the unstructured data are serialized to obtain serialized data;

[0010] The serialized data is processed using a chunking strategy in the graph retrieval enhancement database to obtain a chunked data set. This chunked data set is then subjected to JSON data structuring to obtain JSON data and a training set after JSON data structuring. The training set is used to train, test, and validate the local model. The index value of the JSON data in the graph retrieval enhancement database is also stored in a domestic relational database so that the JSON data can be accessed through the index value in the domestic relational database during the inference phase of the local model.

[0011] The local model is trained using the training set under self-supervised instruction. During the training process, the weight parameters in the pre-trained model file are called to influence the weights of the local model and accelerate the calculation, thereby obtaining a large retrieval language model.

[0012] The retrieval large language model is deployed to the local data center server cluster of the wafer foundry.

[0013] According to the model deployment method for implementing a semiconductor response system as described in the first aspect of this application, the structured data and the unstructured data are serialized to obtain serialized data, including:

[0014] The structured data and the unstructured data are integrated by calling entity extraction methods, semantic tag extraction methods, binary relation extraction methods, multivariate relation extraction methods, or time extraction methods to obtain the serialized data.

[0015] According to the model deployment method for implementing a semiconductor response system as described in the first aspect of this application, the local model includes a convolutional training module and a recursive inference module. Each of the convolutional training module and the recursive inference module includes an embedding layer, a pre-normalization layer, an input projection layer, a main path convolutional layer, a skip connection path projection layer, a main path activation layer, a skip connection activation layer, a gating layer, a selective scan structured state space model, an output projection layer, an output normalization layer, a linear layer, and a softmax layer.

[0016] The selective scanning structured state-space model is implemented using the following algorithm:

[0017] Continuous signaling discrete sequence algorithm, local dependency algorithm, long-range dependency algorithm, structured mask attention mechanism, selective scanning algorithm and dual structured state space algorithm.

[0018] According to the model deployment method for implementing a semiconductor response system as described in the first aspect of this application, the local model is self-supervised trained using the training set, including:

[0019] The training set is tokenized using the embedding layer to obtain a two-dimensional matrix parameter, which includes a sequence length parameter and a channel configuration parameter. The channel configuration parameter includes feature height and feature width. The number of channels in the channel configuration parameter is pre-configured.

[0020] The pre-normalization layer is used to scale and normalize the internal channel parameters obtained based on the configured channel parameters over time to obtain a normalized two-dimensional matrix.

[0021] The normalized two-dimensional matrix is ​​linearly transformed using the input projection layer to obtain the projected two-dimensional matrix.

[0022] The main path convolutional layer is used to perform one-dimensional convolution on the projected two-dimensional matrix along the time dimension channel-by-channel convolutional kernel to obtain a convolutional two-dimensional matrix, which includes time step parameters and weighted channel parameters.

[0023] The activation function of the convolutional two-dimensional matrix is ​​calculated using the main path activation layer to obtain the first activation two-dimensional matrix;

[0024] The activation function of the projected two-dimensional matrix is ​​calculated using the skip connection activation layer to obtain a second activated two-dimensional matrix.

[0025] The gating mechanism utilizes the computational characteristics of the activation function to dynamically filter the gated activation matrix from the first and second activation two-dimensional matrices.

[0026] The selective scanning structured state space model is used to perform convolution or recursive processing on the gated activation two-dimensional matrix to obtain the state two-dimensional matrix;

[0027] The state two-dimensional matrix is ​​convolved using an output projection layer to obtain the two-dimensional matrix at the current time step; or, the state two-dimensional matrix is ​​recursively processed using an output projection layer to obtain the two-dimensional matrix at all time steps.

[0028] The output normalization layer is used to perform root mean square normalization calculation on the two-dimensional matrix of all time steps step by step to obtain a two-dimensional matrix with a fixed scaling distribution.

[0029] A linear layer is used to map the two-dimensional matrix of the fixed scaling distribution into a one-dimensional vector through linear transformation, and a softmax layer is used to map the one-dimensional vector into a probability distribution.

[0030] According to the model deployment method for implementing a semiconductor response system as described in the first aspect of this application, the training set is tokenized using the embedding layer to obtain a two-dimensional matrix parameter including a sequence length parameter and a channel parameter, including:

[0031] The training set is tokenized by the word segmenter in the embedding layer to obtain a discretized token index sequence.

[0032] The token index sequence is transformed into a two-dimensional matrix parameter that includes the sequence length parameter and the configuration channel parameter.

[0033] According to the model deployment method for implementing a semiconductor response system as described in the first aspect of this application, the normalized two-dimensional matrix is ​​linearly transformed using an input projection layer to obtain a projected two-dimensional matrix, including:

[0034] Obtain the pre-configured time steps, training batches, and training generations of the local model;

[0035] The projected two-dimensional matrix is ​​obtained by multiplying the dynamic dimension parameters obtained from the time step, the training batch, and the training generation with the normalized two-dimensional matrix.

[0036] According to the model deployment method for implementing a semiconductor response system described in the first aspect of this application, the main path convolutional layer is used to perform one-dimensional convolution on the projected two-dimensional matrix along the time dimension using channel-by-channel convolutional kernels to obtain a convolutional two-dimensional matrix, including:

[0037] The convolutional two-dimensional matrix is ​​obtained by multiplying the pre-configured one-dimensional matrix in the main path convolutional layer by the projection two-dimensional matrix.

[0038] The two-dimensional convolutional matrix is ​​obtained by performing a weighted double-layer summation along the time dimension channel by channel.

[0039] According to the model deployment method for implementing a semiconductor response system described in the first aspect of this application, the selective scanning structured state-space model is used to perform convolution or recursive processing on the gated activation two-dimensional matrix to obtain the state two-dimensional matrix, including:

[0040] The gated activation two-dimensional matrix is ​​discretized using a continuous signaling discrete sequence algorithm to obtain a continuous signal token sequence and a serialized state space model. The continuous signal token sequence includes a time state representation and step size information. The serialized state space model includes a state input equation, an output equation, and initialization state vector parameters.

[0041] The selective scanning algorithm is used to dynamically calculate the state input equation, the output equation, and the initial state vector parameters step by step over time. In the convolutional training module, the serialized state space model is made to have continuity, and in the recursive inference module, the serialized state space model is made to have selectivity.

[0042] In the convolutional training module, a multi-head parameter set is obtained by dividing the dimensionality of the gated activation two-dimensional matrix. Each head parameter in the multi-head parameter set is a subset of the gated activation two-dimensional matrix. Each head parameter runs a continuous serialized state space model. The continuous serialized state space model is unfolded into a convolutional kernel through a structured mask attention mechanism. A selective scanning structured state space model is obtained through a dual structured state space algorithm. After parallel computation, multiple head parameters are merged to obtain the state two-dimensional matrix containing all continuous time steps.

[0043] In the recursive reasoning module, the gated activation two-dimensional matrix is ​​mapped step-by-step to scalar and state dimension parameters through the first linear transformation algorithm, and each scalar runs a selective serialized state space model; the second linear transformation algorithm maps multiple scalars to obtain a state two-dimensional matrix.

[0044] According to a second aspect of this application, a computing device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.

[0045] According to a third aspect of this application, a computer-readable storage medium is provided that stores a computer program / instructions thereon, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0046] According to a fourth aspect of this application, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the first aspect.

[0047] As will be described in detail below, the model deployment method for implementing a semiconductor response system according to embodiments of this application involves serializing structured and unstructured data obtained locally from the wafer foundry to obtain serialized data. The serialized data is then processed to obtain a training set. This training set is used to perform self-supervised training on a local model using pre-trained model files to obtain a large retrieval language model. This large retrieval language model is then deployed to a local data center server cluster at the wafer foundry. During the local manufacturing process, manufacturing engineers on the production floor can seek solutions from the semiconductor response system deployed in the local data center. Compared to non-AI technologies in traditional wafer manufacturing plants, this reduces the time spent waiting for software engineers, equipment engineers, process engineers, and vendor engineers to arrive on-site, and eliminates the need for significant manpower and resources.

[0048] It should be understood that both the foregoing general description and the following detailed description are exemplary and intended to provide further illustration of the claimed technology. Attached Figure Description

[0049] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0050] Figure 1This is a flowchart illustrating a method for deploying a model of a semiconductor response system according to an embodiment of this application.

[0051] Figure 2 This is a schematic diagram of a local model of an application according to an embodiment of this application.

[0052] Figure 3 This is a flowchart illustrating the self-supervised training of a local model according to an embodiment of this application.

[0053] Figure 4 This is a schematic diagram of a model deployment apparatus for implementing a semiconductor response system according to an embodiment of this application.

[0054] Figure 5 This is a hardware block diagram illustrating a computing device according to an embodiment of this application.

[0055] Figure 6 This is a schematic diagram illustrating a computer-readable storage medium according to an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0057] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0058] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0059] To alleviate the technical problems existing in related technologies, embodiments of this application provide a method, device, and computer program product for deploying a model in a semiconductor response system. This method serializes structured and unstructured data acquired locally from the wafer foundry to obtain serialized data. The serialized data is then processed to obtain a training set. This training set is used to perform self-supervised training on a local model using pre-trained model files, resulting in a large retrieval language model. This large retrieval language model is then deployed to a server cluster in the wafer foundry's local data center. During the local manufacturing process, manufacturing engineers on the production floor can seek solutions from the semiconductor response system deployed in the local data center, encompassing product knowledge, process knowledge, equipment knowledge, testing knowledge, and anomaly handling knowledge. Compared to the non-AI technologies used in traditional wafer manufacturing plants, this reduces the time spent waiting for software engineers, equipment engineers, process engineers, and vendor engineers to arrive on-site, and eliminates the need for significant manpower and resources. Use cases include, but are not limited to: providing maintenance and handling methods for risk prevention and control and clarifying the logic of responsibility division when machine alarms occur, based on historical logs and handling status of machine cases; improving the relevance and interpretability of scheduling and dispatching results based on historical scheduling and dispatching data, thereby providing more granular and flexible scalability solutions; assisting on-duty engineers to quickly sort out problems online and provide relevant logic based on logs, user manuals, and alarm content of various subsystems in the intelligent manufacturing system; and providing relevant engineers with automatic generation of process invoices, report documents, etc., based on wafer fab production rules, information security rules, software coding rules, product testing and shipping rules.

[0060] To facilitate understanding of this embodiment, a detailed description of a model deployment method for implementing a semiconductor response system disclosed in this application embodiment will be provided first. The execution entity of the model deployment method for implementing a semiconductor response system provided in this application embodiment is generally a computing device with certain computing capabilities, such as a terminal device, a server, or other processing devices. In some possible implementations, this model deployment method for implementing a semiconductor response system can be implemented by a processor calling computer-readable instructions stored in memory.

[0061] See Figure 1 The diagram shown is a flowchart of a model deployment method for implementing a semiconductor response system provided in an embodiment of this application. The method includes the following steps:

[0062] Step 101: Obtain large dataset training data for training a general large language model based on a self-attention mechanism architecture.

[0063] Step 102: Train the general large language model and obtain the pre-trained model file.

[0064] Step 103: Obtain structured and unstructured data from the wafer foundry for training and deploying local models.

[0065] Structured data refers to data that strictly adheres to a data model, is easy to search and organize, and is typically stored in relational databases. Structured data has a clear data model and structure, such as a tabular format, making it easy to perform standard queries and analysis. Structured data is usually highly organized and formatted. In the semiconductor field, structured data can include serialized text data. For example, structured data can be descriptive text describing information in an image of wafer defects, or labels used to characterize the type of wafer defects in the image, etc.

[0066] Unstructured data lacks a predefined data model and is typically text or multimedia content. It lacks a fixed format or structure, making it difficult to process with traditional databases and data analysis tools. It usually requires advanced processing methods such as natural language processing or image analysis. Examples of unstructured data include social media posts, video and audio files, documents, and PDF files.

[0067] In one or more embodiments, including but not limited to calling entity extraction methods, semantic tag extraction methods, binary relation extraction methods, multivariate relation extraction methods, or time extraction methods, the structured data and the unstructured data are integrated to obtain the serialized data.

[0068] Step 104: Perform serialization processing on structured and unstructured data to obtain serialized data.

[0069] Step 105: Process the serialized data in the graph retrieval augmentation database using a chunking strategy to obtain a chunked data set, and perform JSON data structuring on the chunked data set to obtain JSON data and a training set after JSON data structuring; wherein, the training set is used to train, test and validate the local model, and the index value of the JSON data in the graph retrieval augmentation database is also stored in a domestic relational database so that the JSON data can be called through the index value of the domestic relational database during the inference stage of the local model.

[0070] Blocking strategies include, but are not limited to, implementations based on structured state-space duality algorithms.

[0071] Step 106: Use the training set to perform self-supervised training on the local model. During the training process, call the weight parameters in the pre-trained model file to influence the weights of the local model and accelerate the calculation, so as to obtain the retrieval large language model.

[0072] Please refer to Figure 2 , Figure 2 This is a schematic diagram of the local model given in this embodiment. Figure 2As shown, the local model includes a convolutional training module and a recursive inference module. Each module of the convolutional training module and the recursive inference module includes an embedding layer, a pre-normalization layer, an input projection layer, a main path convolutional layer, a skip connection path projection layer, a main path activation layer, a skip connection activation layer, a gating layer, a selective scan structured state space model, an output projection layer, an output normalization layer, a linear layer, and a softmax layer.

[0073] The selective scanning structured state-space model is implemented using the following algorithm:

[0074] Continuous signaling discrete sequence algorithm, local dependency algorithm, long-range dependency algorithm, structured mask attention mechanism, selective scanning algorithm and dual structured state space algorithm.

[0075] Based on the specific structure of the local model described above, in some embodiments, such as Figure 3 As shown, step 106 may include the following steps:

[0076] Step 301: Tokenize the training set using an embedding layer to obtain two-dimensional matrix parameters. The two-dimensional matrix parameters include sequence length parameters and channel configuration parameters. The channel configuration parameters include feature height and feature width. The number of channels in the channel configuration parameters is pre-configured.

[0077] In some embodiments, step 301 may include the following steps:

[0078] The training set is tokenized by the word segmenter in the embedding layer to obtain a discretized token index sequence.

[0079] The token index sequence is transformed into a two-dimensional matrix parameter that includes the sequence length parameter and the configuration channel parameter.

[0080] The embedding layer function can be called to process the token index sequence into a two-dimensional input matrix parameter X. (L,D) Where X represents the discretized token index sequence after matrix transformation, L represents the sequence length parameter, and D represents the configuration channel parameter of the model configuration, which includes feature height and feature width.

[0081] Step 302: Use a pre-normalization layer to scale and normalize the internal channel parameters obtained based on the configured channel parameters over time to obtain a normalized two-dimensional matrix.

[0082] In this embodiment, a normalized two-dimensional matrix with a fixed scaling distribution can be calculated by calling the root mean square normalization function based on the two-dimensional input matrix parameters. The root mean square normalization function normalizes the learnable parameters by scaling them at each time step using the feature channels, feature height, and feature width as dimensions. This ensures that the parameters are learnable with each time step while maintaining a constant dimension of the two-dimensional matrix parameters, and is unaffected by the variability of the sequence length parameter. It also eliminates parameter shift calculations to accelerate model computation at this layer and prevents ICS (Internal Covariate Shift).

[0083] Step 303: Perform a linear transformation on the normalized two-dimensional matrix using the input projection layer to obtain the projected two-dimensional matrix.

[0084] In some embodiments, step 303 may include the following steps:

[0085] Obtain the pre-configured time steps, training batches, and training generations of the local model;

[0086] The projected two-dimensional matrix is ​​obtained by multiplying the dynamic dimension parameters obtained from the time step, the training batch, and the training generation with the normalized two-dimensional matrix.

[0087] Based on the normalized two-dimensional matrix, a linear transformation is invoked to obtain the projected two-dimensional matrix. The product calculation uses the following formula: X proj,t This represents the projected two-dimensional matrix, where proj is the abbreviation for projection, t represents the current time step, and X′ is the projection matrix. t This represents the projected two-dimensional matrix of the previous time step, and the standard matrix at t=0. D represents a normalized two-dimensional matrix, and D represents the model configuration channel parameters. inner This represents the dynamic internal channel parameters acquired by the model during training, based on the time steps, training batches, and training generations pre-configured by the local model. (D) inner ≤D.

[0088] Step 304: Use the main path convolutional layer to perform one-dimensional convolution on the projected two-dimensional matrix along the time dimension channel by channel convolutional kernel to obtain a convolutional two-dimensional matrix, which includes time step parameters and weighted channel parameters.

[0089] In some embodiments, step 304 may include the following steps:

[0090] The convolutional two-dimensional matrix is ​​obtained by multiplying the pre-configured one-dimensional matrix in the main path convolutional layer by the projection two-dimensional matrix.

[0091] The two-dimensional convolutional matrix is ​​obtained by performing a weighted double-layer summation along the time dimension channel by channel.

[0092] Based on the projected two-dimensional matrix, a one-dimensional convolution function is invoked to calculate the one-dimensional convolution along the time dimension, channel-by-channel, and then weighted double-layer summation is performed to obtain the convolutional two-dimensional matrix. The formula used for the weighted double-layer summation is as follows: L represents the sequence length parameter, D represents the channel parameters configured in the model, conv stands for convolution, k represents the total number of time steps, i represents the index of the time step (0 ≤ i ≤ k-1), c represents the index of the input channel parameters (0 ≤ c ≤ k-1), j represents the index of the output channel parameters (0 ≤ j ≤ D), and W... i,c,j This represents the convolution kernel parameters, specifically the parameters of the convolution kernel at time step i, the c-th input channel, and the j-th output channel, summed within the inner layer. This indicates iterating through each channel and summing the results, with the outermost layer being summed. This indicates summation over time steps. This represents a two-dimensional convolutional matrix, specifically the two-dimensional convolutional matrix of the i-th input channel and the j-th output channel.

[0093] By leveraging the computational characteristics of one-dimensional convolution, when j>c during inner-layer summation, the local model amplifies the learning of feature channel parameters and conversely reduces the learning of feature channel parameters, thus achieving feature scaling. Outer-layer summation achieves feature translation along the time step, and double-layer weighted summation implements a local dependency algorithm, enabling the model to capture rich contextual information.

[0094] Step 305: Calculate the activation function of the convolutional two-dimensional matrix using the main path activation layer to obtain the first activated two-dimensional matrix.

[0095] Based on the convolutional two-dimensional matrix, the SiLU (Sigmoid Linear Unit) activation function is called to calculate and obtain the first activation two-dimensional matrix. The calculation formula used for the activation function is as follows: "act" is the abbreviation for activation, "conv" is the abbreviation for convolution, "L" represents the sequence length parameter, "D" represents the channel parameter in the model configuration, "t" represents the time step index (0 ≤ t ≤ L-1), "d" represents the channel index (0 ≤ d ≤ D-1), and the sigma function. This represents the first activation two-dimensional matrix, that is, the first activation two-dimensional matrix of the d-th channel at the t-th time step.

[0096] By leveraging the characteristics of mathematical computation, nonlinear transformations are performed on each sequence length parameter and channel parameter of the convolutional two-dimensional matrix step by step and channel by channel, which makes up for the lack of prior knowledge in the convolutional layer and enhances the model's ability to learn complex knowledge.

[0097] Step 306: Calculate the activation function of the projected two-dimensional matrix using a skip connection activation layer to obtain the second activated two-dimensional matrix.

[0098] Based on the projected two-dimensional matrix, the SiLU activation function is called to calculate and obtain the second activation two-dimensional matrix. The calculation formula for the SiLU activation function is as follows: `skip` is an abbreviation for skip connection, `proj` is an abbreviation for projection, `L` represents the sequence length parameter, `D` represents the channel parameter of the model configuration, `t` represents the time step index (0 ≤ t ≤ L-1), `d` represents the channel index (0 ≤ d ≤ D-1), and the sigma function. This represents the second activation two-dimensional matrix, which is the first activation two-dimensional matrix of the d-th channel at the t-th time step.

[0099] By leveraging the characteristics of mathematical computation, nonlinear transformations are performed on each sequence length parameter and channel parameter of the convolutional two-dimensional matrix step by step and channel by channel, which makes up for the lack of prior knowledge in the convolutional layer and enhances the model's ability to learn complex knowledge.

[0100] Step 307: Using gating to dynamically select a gating activation matrix from the first and second activation two-dimensional matrices by utilizing the computational characteristics of the activation function.

[0101] Based on the first and second activation two-dimensional matrices, a gated activation two-dimensional matrix is ​​dynamically selected using a matrix multiplication function. The formula for the matrix multiplication function is as follows: "Gated" is the abbreviation for gated, "act" is the abbreviation for activation, "skip" is the abbreviation for skip connection, "L" represents the sequence length parameter, "D" represents the channel parameter configured in the model, "t" represents the time step index (0 ≤ t ≤ L-1), and "d" represents the channel index (0 ≤ d ≤ D-1). This represents the gating activation matrix parameters, specifically the gating activation matrix parameters for the d-th channel at the t-th time step.

[0102] By leveraging the computational properties of activation functions and matrix multiplication, when the parameters of the first activation two-dimensional matrix are approximately zero and less than those of the second activation two-dimensional matrix, the feature information of the main path activation layer is suppressed, and the feature information of the skip connection layer is used; otherwise, the feature information of the main path activation layer is used. This method implements the local dependency algorithm and the long-distance dependency algorithm, enabling the model to obtain not only local dependencies (short-term memory) through the main path convolutional layer, but also long-distance dependencies (long-term memory) through the skip connection layer. It addresses the forgetting nature of large language models as the time step t increases with the number of response rounds, and through the gating property, allows the model to achieve the longest possible memory under current hardware conditions.

[0103] Step 308: Use a selective scanning structured state space model to perform convolution or recursive processing on the gated activation two-dimensional matrix to obtain the state two-dimensional matrix.

[0104] In some embodiments, step 308 may include the following steps:

[0105] The gated activation two-dimensional matrix is ​​discretized using a continuous signaling discrete sequence algorithm to obtain a continuous signal token sequence and a serialized state space model. The continuous signal token sequence includes a time state representation and step size information. The serialized state space model includes a state input equation, an output equation, and initialization state vector parameters.

[0106] The selective scanning algorithm is used to dynamically calculate the state input equation, the output equation, and the initial state vector parameters step by step over time. In the convolutional training module, the serialized state space model is made to have continuity, and in the recursive inference module, the serialized state space model is made to have selectivity.

[0107] In the convolutional training module, a multi-head parameter set is obtained by dividing the dimensionality of the gated activation two-dimensional matrix. Each head parameter in the multi-head parameter set is a subset of the gated activation two-dimensional matrix. Each head parameter runs a continuous serialized state space model. The continuous serialized state space model is unfolded into a convolutional kernel through a structured mask attention mechanism. A selective scanning structured state space model is obtained through a dual structured state space algorithm. After parallel computation, multiple head parameters are merged to obtain the state two-dimensional matrix containing all continuous time steps.

[0108] In the recursive reasoning module, the gated activation two-dimensional matrix is ​​mapped step-by-step to scalar and state dimension parameters through the first linear transformation algorithm, and each scalar runs a selective serialized state space model; the second linear transformation algorithm maps multiple scalars to obtain a state two-dimensional matrix.

[0109] A state two-dimensional matrix is ​​obtained by performing convolution or recursive processing based on a gated activation two-dimensional matrix.

[0110] The continuous signal-to-discrete sequence algorithm is used to discretize the gated activation two-dimensional matrix, obtaining a continuous signal token sequence and a serialized state-space model. The continuous signal token sequence includes a time state representation and step size information, while the serialized state-space model includes a state input equation, an output equation, and initialized state vector parameters, representing a concrete implementation of the state-space model theory. The state input equation is represented as h. t =Ah t-1 +Bt x t The output equation is represented as y t =C t h t +Dx t t represents the time step index, h t This represents the two-dimensional state matrix at the current time step, where A represents the initial state vector parameters, and h... t-1 B represents the result of the input equation calculation at the previous time step. t Indicates dynamic input parameters, y t The output of the current time step represents the state, C. t This indicates dynamic output parameters, and D indicates jump connection parameters.

[0111] In the convolutional training module, a multi-head parameter set is obtained by using a dimension-splitting gating activation two-dimensional matrix, calculated as follows: L represents the sequence length parameter, D represents the channel parameter configured in the model, and H represents the size of the multi-head parameter set. Let represent the gating activation matrix parameters. Each head parameter in the multi-head parameter set is a subset of the gating activation two-dimensional matrix, and each head parameter is represented as . h represents the current state. Each header parameter runs a continuous, serialized state-space model, calculated using the following formula: SSM is an abbreviation for State-Space Model Function. The convolution kernel representation of the current state, This represents the result parameter after convolution of each head parameter. The final convolution result parameter is obtained by merging the result parameters of each head parameter within the range 0 ≤ h ≤ h-1 through a convolution function. The above calculation formula indicates that the initial state vector parameter A, dynamic input parameter B, dynamic output parameter C, and skip connection parameter D will dynamically change with state transitions. In each state, the dynamic input parameter B... t Dynamic output parameter C t It will change dynamically over time.

[0112] The convolutional kernel captures each state h at each time step. t The evolution process, through convolution, enables the local model to learn not only with each time step but also with each state transition. The initialization of the state vector parameters aims to simplify the calculation process at each time step and for each state transition. Taking the input equation as an example, the calculation formula for the continuous, sequential state-space model during the convolution process is as follows: i represents the time step index. When i = t, t = 0, A t-u =A 0=I, where I represents the identity matrix, and when i = t-1, t-1 = 1, A t-i =A 1 =A, when i = t-2, t-1 = 2, A t-i =A 2 As demonstrated by the deduction process, keeping the initial state vector parameters constant at each time step reduces computation. During each state transition, the parameters decay or increase with each state transition through exponential changes, achieving the aforementioned continuity and computational efficiency. The convolution process, by strengthening the local and long-range dependency algorithms, enables the model to capture not only short-term memories but also long-term memories during state transitions.

[0113] The continuous, sequential state-space model is unfolded into a convolutional kernel using a structured mask attention mechanism, calculated as follows: When i ≤ t, C t A t-i B i Maintain dynamic calculation; when i > t, C t A t-i B i The value of 0 indicates a dynamic change over time, allowing the model to dynamically select the current time step for either local or long-distance dependency, thus achieving selectivity.

[0114] The dual structured state space algorithm utilizes the efficiency of semi-separable matrix computation to selectively scan the structured state space model. After parallel computation, multiple head parameters are merged to obtain the two-dimensional state matrix containing all continuous time steps.

[0115] In the recursive inference module, the gated activation two-dimensional matrix is ​​mapped step-by-step to scalar and state dimension parameters using the first linear transformation algorithm. The input calculation formula is as follows: The output calculation formula is as follows: N represents the number of recursions required after converting parallel convolution to serial computation. Each scalar runs one of the selectively serialized state-space models. Multiple scalars are mapped to a two-dimensional state matrix using a second linear transformation algorithm. Both the first and second linear transformation algorithms can be split and merged by calling matrix multiplication functions.

[0116] The time complexity of the above convolutional training module is O(LDH), where H represents the size of the multi-head set, and the time complexity of the recursive inference module is O(LDn), where n represents the number of recursions.

[0117] Furthermore, according to the model deployment method for implementing a semiconductor response system according to the third aspect of this application, the output projection layer is employed, including:

[0118] Convolution processing is performed on the state two-dimensional matrix to obtain the two-dimensional matrix of the current time step; or, the state two-dimensional matrix is ​​recursively processed by the output projection layer to obtain the two-dimensional matrix of all time steps.

[0119] Step 309: Perform convolution on the state two-dimensional matrix using the output projection layer to obtain the two-dimensional matrix of the current time step; or, perform recursive processing on the state two-dimensional matrix using the output projection layer to obtain the two-dimensional matrix of all time steps.

[0120] Step 310: Use the output normalization layer to perform root mean square normalization calculation on the two-dimensional matrix of all time steps step by step to obtain a two-dimensional matrix with a fixed scaling distribution.

[0121] In some embodiments, step 310 may include the following steps:

[0122] The root mean square normalization calculation is performed step-by-step on the two-dimensional matrix based on all time steps to obtain a two-dimensional matrix with a fixed scaling distribution.

[0123] Furthermore, according to the model deployment method for implementing a semiconductor response system according to the third aspect of this application, the linear layer and softmax layer are employed, including:

[0124] A two-dimensional matrix based on the fixed scaling distribution is mapped to a one-dimensional vector through a linear transformation;

[0125] The one-dimensional tensor is mapped to a probability distribution by calling the softmax function.

[0126] Step 311: Use a linear layer to map a two-dimensional matrix with a fixed scaling distribution into a one-dimensional vector through linear transformation, and use a softmax layer to map the one-dimensional vector into a probability distribution.

[0127] Step 107: Deploy the large language model to the local data center server cluster of the wafer foundry.

[0128] The solution provided in this embodiment serializes structured and unstructured data acquired locally from the wafer foundry to obtain serialized data. This serialized data is then processed to obtain a training set. The training set, along with pre-trained model files, is used to perform self-supervised training on the local model, resulting in a large retrieval language model. This large retrieval language model is then deployed to the wafer foundry's local data center server cluster. During the local manufacturing process, manufacturing engineers on the production floor can seek solutions from the semiconductor response system deployed in the local data center. Compared to the non-AI technologies used in traditional wafer fabrication plants, this reduces the time spent waiting for software engineers, equipment engineers, process engineers, and vendor engineers to arrive on-site, and eliminates the need for significant manpower and resources.

[0129] This application also provides a model deployment apparatus for implementing a semiconductor response system, which is used to execute the model deployment method for implementing a semiconductor response system provided in any of the above embodiments. Figure 4 As shown, the device includes:

[0130] The first acquisition module 41 is used to acquire big data training set data for training a general large language model based on a self-attention mechanism architecture;

[0131] Training module 42 is used to train the general large language model and obtain a pre-trained model file;

[0132] The second acquisition module 43 is used to acquire structured and unstructured data from the wafer foundry for training and deploying local models.

[0133] The first processing module 44 is used to perform serialization processing on the structured data and the unstructured data to obtain serialized data;

[0134] The second processing module 45 is used to process the serialized data in the graph retrieval enhancement database using a block strategy to obtain a block data set, and to perform JSON data structuring on the block data set to obtain JSON data and a training set after JSON data structuring; wherein, the training set is used to train, test and validate the local model, and the index value of the JSON data in the graph retrieval enhancement database is also stored in a domestic relational database so that the JSON data can be called through the index value in the domestic relational database during the inference stage of the local model;

[0135] Training module 46 is used to perform self-supervised training on the local model using the training set. During the training process, the weight parameters in the pre-trained model file are called to influence the weights of the local model and accelerate the calculation, so as to obtain a large retrieval language model.

[0136] Deployment module 47 is used to deploy the retrieval large language model to the local data center server cluster of the wafer foundry.

[0137] In one or more embodiments, the first processing module 44 is used to:

[0138] The structured data and the unstructured data are integrated by calling entity extraction methods, semantic tag extraction methods, binary relation extraction methods, multivariate relation extraction methods, or time extraction methods to obtain the serialized data.

[0139] In one or more embodiments, the local model includes a convolutional training module and a recursive inference module, each of which includes an embedding layer, a pre-normalization layer, an input projection layer, a main path convolutional layer, a skip connection path projection layer, a main path activation layer, a skip connection activation layer, a gating layer, a selective scan structured state space model, an output projection layer, an output normalization layer, a linear layer, and a softmax layer.

[0140] The selective scanning structured state-space model is implemented using the following algorithm:

[0141] Continuous signaling discrete sequence algorithm, local dependency algorithm, long-range dependency algorithm, structured mask attention mechanism, selective scanning algorithm and dual structured state space algorithm.

[0142] In one or more embodiments, the training module 46 is used for:

[0143] The training set is tokenized using the embedding layer to obtain a two-dimensional matrix parameter, which includes a sequence length parameter and a channel configuration parameter. The channel configuration parameter includes feature height and feature width. The number of channels in the channel configuration parameter is pre-configured.

[0144] The pre-normalization layer is used to scale and normalize the internal channel parameters obtained based on the configured channel parameters over time to obtain a normalized two-dimensional matrix.

[0145] The normalized two-dimensional matrix is ​​linearly transformed using the input projection layer to obtain the projected two-dimensional matrix.

[0146] The main path convolutional layer is used to perform one-dimensional convolution on the projected two-dimensional matrix along the time dimension channel-by-channel convolutional kernel to obtain a convolutional two-dimensional matrix, which includes time step parameters and weighted channel parameters.

[0147] The activation function of the convolutional two-dimensional matrix is ​​calculated using the main path activation layer to obtain the first activation two-dimensional matrix;

[0148] The activation function of the projected two-dimensional matrix is ​​calculated using the skip connection activation layer to obtain a second activated two-dimensional matrix.

[0149] The gating mechanism utilizes the computational characteristics of the activation function to dynamically filter the gated activation matrix from the first and second activation two-dimensional matrices.

[0150] The selective scanning structured state space model is used to perform convolution or recursive processing on the gated activation two-dimensional matrix to obtain the state two-dimensional matrix;

[0151] The state two-dimensional matrix is ​​convolved using an output projection layer to obtain the two-dimensional matrix at the current time step; or, the state two-dimensional matrix is ​​recursively processed using an output projection layer to obtain the two-dimensional matrix at all time steps.

[0152] The output normalization layer is used to perform root mean square normalization calculation on the two-dimensional matrix of all time steps step by step to obtain a two-dimensional matrix with a fixed scaling distribution.

[0153] A linear layer is used to map the two-dimensional matrix of the fixed scaling distribution into a one-dimensional vector through linear transformation, and a softmax layer is used to map the one-dimensional vector into a probability distribution.

[0154] In one or more embodiments, the training module 46 is used for:

[0155] The training set is tokenized by the word segmenter in the embedding layer to obtain a discretized token index sequence.

[0156] The token index sequence is transformed into a two-dimensional matrix parameter that includes the sequence length parameter and the configuration channel parameter.

[0157] In one or more embodiments, the training module 46 is used for:

[0158] Obtain the pre-configured time steps, training batches, and training generations of the local model;

[0159] The projected two-dimensional matrix is ​​obtained by multiplying the dynamic dimension parameters obtained from the time step, the training batch, and the training generation with the normalized two-dimensional matrix.

[0160] In one or more embodiments, the training module 46 is used for:

[0161] The convolutional two-dimensional matrix is ​​obtained by multiplying the pre-configured one-dimensional matrix in the main path convolutional layer by the projection two-dimensional matrix.

[0162] The two-dimensional convolutional matrix is ​​obtained by performing a weighted double-layer summation along the time dimension channel by channel.

[0163] In one or more embodiments, the training module 46 is used for:

[0164] The gated activation two-dimensional matrix is ​​discretized using a continuous signaling discrete sequence algorithm to obtain a continuous signal token sequence and a serialized state space model. The continuous signal token sequence includes a time state representation and step size information. The serialized state space model includes a state input equation, an output equation, and initialization state vector parameters.

[0165] The selective scanning algorithm is used to dynamically calculate the state input equation, the output equation, and the initial state vector parameters step by step over time. In the convolutional training module, the serialized state space model is made to have continuity, and in the recursive inference module, the serialized state space model is made to have selectivity.

[0166] In the convolutional training module, a multi-head parameter set is obtained by dividing the dimensionality of the gated activation two-dimensional matrix. Each head parameter in the multi-head parameter set is a subset of the gated activation two-dimensional matrix. Each head parameter runs a continuous serialized state space model. The continuous serialized state space model is unfolded into a convolutional kernel through a structured mask attention mechanism. A selective scanning structured state space model is obtained through a dual structured state space algorithm. After parallel computation, multiple head parameters are merged to obtain the state two-dimensional matrix containing all continuous time steps.

[0167] In the recursive reasoning module, the gated activation two-dimensional matrix is ​​mapped step-by-step to scalar and state dimension parameters through the first linear transformation algorithm, and each scalar runs a selective serialized state space model; the second linear transformation algorithm maps multiple scalars to obtain a state two-dimensional matrix.

[0168] The model deployment apparatus for implementing a semiconductor response system provided in this application embodiment and the model deployment method for implementing a semiconductor response system provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0169] This application also provides a computing device for executing the above-described method for deploying a model of a semiconductor response system. Please refer to... Figure 5 It illustrates a schematic diagram of a computing device provided by some embodiments of this application. For example... Figure 5 As shown, the computing device 5 includes a processor 501, a memory 502, and a bus 503. The processor 501 and the memory 502 are connected via the bus 503. The memory 502 stores a computer program that can run on the processor 501. When the processor 501 runs the computer program, it executes the model deployment method for implementing a semiconductor response system provided in any of the foregoing embodiments of this application.

[0170] The memory 502 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device.

[0171] Bus 503 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. Memory 502 is used to store programs. After receiving an execution instruction, processor 501 executes the program. The model deployment method for implementing a semiconductor response system disclosed in any of the foregoing embodiments of this application can be applied to processor 501, or implemented by processor 501.

[0172] Processor 501 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 501 or by instructions in software form. Processor 501 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), an Off-the-shelf Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 502. Processor 501 reads the information in memory 502 and, in conjunction with its hardware, completes the steps of the above method.

[0173] The computing device provided in this application embodiment and the model deployment method for implementing a semiconductor response system provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0174] This application also provides a computer-readable storage medium corresponding to the model deployment method for implementing a semiconductor response system provided in the foregoing embodiments. Please refer to... Figure 6 The computer-readable storage medium shown is an optical disc 60, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it executes the model deployment method for implementing a semiconductor response system provided in any of the foregoing embodiments.

[0175] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0176] The computer-readable storage medium provided in the above embodiments of this application and the model deployment method for implementing a semiconductor response system provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0177] It should be noted that:

[0178] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0179] Similarly, it should be understood that, for the sake of brevity and to aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of this application, various features of this application are sometimes grouped together in a single embodiment, figure, or description thereof. However, this disclosure should not be construed as reflecting a schematic diagram in which the claimed application requires more features than expressly recited in each claim. Rather, as reflected in the following claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0180] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0181] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for deploying a model of a semiconductor response system, characterized in that, include: Obtain large datasets for training a general-purpose large language model based on a self-attention mechanism architecture; Train the general large language model and obtain the pre-trained model file; Obtain structured and unstructured data locally from the wafer foundry for training and deploying local models; The structured data and the unstructured data are serialized to obtain serialized data; The serialized data is processed using a chunking strategy in the graph retrieval enhancement database to obtain a chunked data set. This chunked data set is then subjected to JSON data structuring to obtain JSON data and a training set after JSON data structuring. The training set is used to train, test, and validate the local model. The index value of the JSON data in the graph retrieval enhancement database is also stored in a domestic relational database so that the JSON data can be accessed through the index value in the domestic relational database during the inference phase of the local model. The local model is trained using the training set under self-supervised instruction. During the training process, the weight parameters in the pre-trained model file are called to influence the weights of the local model and accelerate the calculation, thereby obtaining a large retrieval language model. The retrieval large language model is deployed to the local data center server cluster of the wafer foundry.

2. The method according to claim 1, characterized in that, The structured data and the unstructured data are serialized to obtain serialized data, including: The structured data and the unstructured data are integrated by calling entity extraction methods, semantic tag extraction methods, binary relation extraction methods, multivariate relation extraction methods, or time extraction methods to obtain the serialized data.

3. The method according to claim 1, characterized in that, The local model includes a convolutional training module and a recursive inference module. Each of the convolutional training module and the recursive inference module includes an embedding layer, a pre-normalization layer, an input projection layer, a main path convolutional layer, a skip connection path projection layer, a main path activation layer, a skip connection activation layer, a gating layer, a selective scan structured state space model, an output projection layer, an output normalization layer, a linear layer, and a softmax layer. The selective scanning structured state-space model is implemented using the following algorithm: Continuous signaling discrete sequence algorithm, local dependency algorithm, long-range dependency algorithm, structured mask attention mechanism, selective scanning algorithm and dual structured state space algorithm.

4. The method according to claim 3, characterized in that, Self-supervised training of the local model using the training set includes: The training set is tokenized using the embedding layer to obtain a two-dimensional matrix parameter, which includes a sequence length parameter and a channel configuration parameter. The channel configuration parameter includes feature height and feature width. The number of channels in the channel configuration parameter is pre-configured. The pre-normalization layer is used to scale and normalize the internal channel parameters obtained based on the configured channel parameters over time to obtain a normalized two-dimensional matrix. The normalized two-dimensional matrix is ​​linearly transformed using the input projection layer to obtain the projected two-dimensional matrix. The main path convolutional layer is used to perform one-dimensional convolution on the projected two-dimensional matrix along the time dimension channel-by-channel convolutional kernel to obtain a convolutional two-dimensional matrix, which includes time step parameters and weighted channel parameters. The activation function of the convolutional two-dimensional matrix is ​​calculated using the main path activation layer to obtain the first activation two-dimensional matrix; The activation function of the projected two-dimensional matrix is ​​calculated using the skip connection activation layer to obtain a second activated two-dimensional matrix. The gating mechanism utilizes the computational characteristics of the activation function to dynamically filter the gated activation matrix from the first and second activation two-dimensional matrices. The selective scanning structured state space model is used to perform convolution or recursive processing on the gated activation two-dimensional matrix to obtain the state two-dimensional matrix; The state two-dimensional matrix is ​​convolved using an output projection layer to obtain the two-dimensional matrix at the current time step; or, the state two-dimensional matrix is ​​recursively processed using an output projection layer to obtain the two-dimensional matrix at all time steps. The output normalization layer is used to perform root mean square normalization calculation on the two-dimensional matrix of all time steps step by step to obtain a two-dimensional matrix with a fixed scaling distribution. A linear layer is used to map the two-dimensional matrix of the fixed scaling distribution into a one-dimensional vector through linear transformation, and a softmax layer is used to map the one-dimensional vector into a probability distribution.

5. The method according to claim 4, characterized in that, The training set is tokenized using the embedding layer to obtain a two-dimensional matrix parameter including sequence length and channel parameters, including: The training set is tokenized by the word segmenter in the embedding layer to obtain a discretized token index sequence. The token index sequence is transformed into a two-dimensional matrix parameter that includes the sequence length parameter and the configuration channel parameter.

6. The method according to claim 4, characterized in that, The normalized two-dimensional matrix is ​​linearly transformed using an input projection layer to obtain a projected two-dimensional matrix, including: Obtain the pre-configured time steps, training batches, and training generations of the local model; The projected two-dimensional matrix is ​​obtained by multiplying the dynamic dimension parameters obtained from the time step, the training batch, and the training generation with the normalized two-dimensional matrix.

7. The method according to claim 4, characterized in that, The main path convolutional layer is used to perform one-dimensional convolution on the projected two-dimensional matrix along the time dimension using channel-by-channel convolutional kernels to obtain a convolutional two-dimensional matrix, including: The convolutional two-dimensional matrix is ​​obtained by multiplying the pre-configured one-dimensional matrix in the main path convolutional layer by the projection two-dimensional matrix. The two-dimensional convolutional matrix is ​​obtained by performing a weighted double-layer summation along the time dimension channel by channel.

8. The method according to claim 4, characterized in that, The selective scanning structured state-space model is used to perform convolution or recursive processing on the gated activation two-dimensional matrix to obtain the state two-dimensional matrix, including: The gated activation two-dimensional matrix is ​​discretized using a continuous signaling discrete sequence algorithm to obtain a continuous signal token sequence and a serialized state space model. The continuous signal token sequence includes a time state representation and step size information. The serialized state space model includes a state input equation, an output equation, and initialization state vector parameters. The selective scanning algorithm is used to dynamically calculate the state input equation, the output equation, and the initial state vector parameters step by step over time. In the convolutional training module, the serialized state space model is made to have continuity, and in the recursive inference module, the serialized state space model is made to have selectivity. In the convolutional training module, a multi-head parameter set is obtained by dividing the dimensionality of the gated activation two-dimensional matrix. Each head parameter in the multi-head parameter set is a subset of the gated activation two-dimensional matrix. Each head parameter runs a continuous serialized state space model. The continuous serialized state space model is unfolded into a convolutional kernel through a structured mask attention mechanism. A selective scanning structured state space model is obtained through a dual structured state space algorithm. After parallel computation, multiple head parameters are merged to obtain the state two-dimensional matrix containing all continuous time steps. In the recursive reasoning module, the gated activation two-dimensional matrix is ​​mapped step-by-step to scalar and state dimension parameters through the first linear transformation algorithm, and each scalar runs a selective serialized state space model; the second linear transformation algorithm maps multiple scalars to obtain a state two-dimensional matrix.

9. A computing device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-8.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-8.

Images