Large model precision overflow processing method and device based on numerical scaling and medium
By numerical scaling of the input data of the quantized large model, it is lower than the maximum representation ability of the model, and numerical restoration is performed in the output data, the accuracy overflow problem of the quantitative large model when dealing with extreme numerical values is solved, and the accuracy and stability are improved, while maintaining the lightweight characteristics of the model.
Patent Information
- Application Number
- CN202510050552.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-13
- Publication Date
- 2025-04-18
AI Technical Summary
Quantitative large models are prone to cause precision overflow when processing complex contexts containing extreme values, resulting in distortion of model inference results.
By obtaining the input data of the quantitative large model, determining the maximum actual numerical value and maximum representation ability, generating a scaling factor, scaling the input data to make it lower than the maximum representation ability, and numerical restoration is performed in the output data.
It effectively solves the accuracy overflow problem of quantitative large models when dealing with large numerical contexts, ensures the accuracy of inference results and the stability of model operation, and maintains the lightweight advantages of quantitative models.
Smart Images
Figure CN119848513A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computers, and in particular to a method, device and medium for processing large model precision overflow based on numerical scaling. Background Art
[0002] With the development of deep learning technology, quantizing large models has become an important means to reduce computing resource consumption and accelerate reasoning speed. However, the precision loss in the quantization process, especially when dealing with complex contexts containing extreme values, can easily lead to precision overflow, resulting in distortion of model reasoning results and affecting the final application effect.
[0003] The traditional solution is to directly expand the numerical representation range of the model. Although this can alleviate the problem to a certain extent, it will increase the storage and calculation overhead of the model, which goes against the original intention of quantization. Summary of the invention
[0004] In order to solve the above problems, this application proposes a large model precision overflow processing method based on numerical scaling, including:
[0005] Obtain input data of the current quantitative large model, and determine the maximum actual value thereof based on the input data;
[0006] Determining the maximum representation capability of the quantized large model, and generating a corresponding scaling factor according to the maximum actual value and the maximum representation capability;
[0007] Scaling the input data by the scaling factor so that the scaled input data is lower than the maximum capability representation;
[0008] The scaled input data is input into the quantized large model, and the output data of the quantized large model is obtained, and the output data is numerically restored according to the scaling factor.
[0009] In one example, after scaling the input data by the scaling factor, the method further includes:
[0010] A numerical mapping table is established to facilitate numerical restoration of the output data according to the numerical mapping table; the numerical mapping table contains original numerical values corresponding to the input data and the corresponding scaled numerical values after scaling.
[0011] In one example, generating a corresponding scaling factor according to the maximum actual value and the maximum representation capability specifically includes:
[0012] If the maximum actual value is higher than the maximum representation capability, the corresponding proportionality coefficient is obtained by dividing the maximum representation capability and the maximum actual value;
[0013] Determining a correction coefficient according to the model depth of the quantitative large model;
[0014] The proportional coefficient is corrected by the correction coefficient so that the proportional coefficient is reduced to obtain a corresponding scaling factor; wherein the lower the value of the scaling factor, the higher the scaling capability.
[0015] In one example, determining the correction coefficient according to the model depth of the quantized large model specifically includes:
[0016] Obtaining a model depth according to the number of intermediate layers of the quantized large model;
[0017] If the model depth is lower than a preset depth, determining the correction coefficient to be a preset first fixed value;
[0018] If the model depth is higher than a preset depth, the preset first fixed value is correspondingly reduced according to the degree to which the model depth is higher than the preset depth to obtain a correction coefficient.
[0019] In one example, generating a corresponding scaling factor according to the maximum actual value and the maximum representation capability specifically includes:
[0020] If the maximum actual value is lower than the maximum representation capability, a preset second fixed value is used as the scaling factor.
[0021] In one example, scaling the input data by the scaling factor specifically includes:
[0022] Acquire key data and non-key data in the input data; wherein the key data and non-key data are determined by data tags of the input data;
[0023] The non-critical data is scaled by the scaling factor, and the critical data is not scaled and directly input into the quantitative large model.
[0024] In one example, the method further includes:
[0025] For each input data, determine the data source of the input data and determine the reliability of the data source; and for the business type corresponding to the input data, determine the business relevance of the input data; and determine the uniqueness of the input data based on the data difference between the input data and other input data;
[0026] Determine whether the input data is critical data based on the reliability, the business relevance, and the uniqueness.
[0027] In one example, scaling the input data by the scaling factor specifically includes:
[0028] When the input data is text data, converting the text data into a word vector and scaling a first preset dimension of the word vector;
[0029] When the input data is audio data, for each sampling point of the audio data, scaling the second preset dimension of the sampling point;
[0030] When the input data is image data, for each pixel value in the image data, the third preset dimension of the pixel value is scaled.
[0031] On the other hand, the present application also proposes a large model precision overflow processing device based on numerical scaling, comprising:
[0032] at least one processor; and,
[0033] a memory communicatively connected to the at least one processor; wherein,
[0034] The memory stores instructions that can be executed 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 execute a large model precision overflow processing method based on numerical scaling as described in any of the above examples.
[0035] On the other hand, the present application also proposes a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to be: a large model precision overflow processing method based on numerical scaling as described in any of the above examples.
[0036] The large model precision overflow processing method based on numerical scaling proposed in this application can bring the following results:
[0037] Beneficial effects:
[0038] 1. Improve accuracy and stability: Through the numerical scaling strategy, the precision overflow problem of large quantitative models when processing large numerical contexts is effectively solved, ensuring the accuracy of inference results and the stability of model operation.
[0039] 2. Resource efficiency: Compared with the method of directly expanding the model representation range, it achieves effective processing of large values without significantly increasing the model calculation and storage burden, maintaining the lightweight advantage of the quantitative model.
[0040] 3. Versatility and flexibility: It is highly versatile and suitable for various types of large quantitative models and different application scenarios. At the same time, the selection of scaling factors provides flexibility for different accuracy requirements.
[0041] 4. Easy to implement and integrate: With clear pre-processing, inference monitoring, and post-processing steps, the solution is easy to implement and integrate with existing quantitative model frameworks and inference engines. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0043] Figure 1 It is a flow chart of a method for processing large model precision overflow based on numerical scaling in an embodiment of the present application;
[0044] Figure 2 This is a flow chart of a method for processing large model precision overflow based on numerical scaling in one scenario in an embodiment of the present application;
[0045] Figure 3 It is a schematic diagram of a large model precision overflow processing device based on numerical scaling in an embodiment of the present application. DETAILED DESCRIPTION
[0046] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.
[0047] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.
[0048] like Figure 1 As shown, the embodiment of the present application provides a large model precision overflow processing method based on numerical scaling, including:
[0049] S101: Obtain input data of the current large quantitative model, and determine the maximum actual value thereof based on the input data.
[0050] The quantization process is a technology that optimizes large models by converting the data representation in the model from high precision to low precision. It can convert the parameters and calculations originally represented by high-precision data types (such as 32-bit floating point numbers) in the large model into low-precision data types (such as 8-bit integers, 16-bit floating point numbers, etc.) for representation and processing. A quantized large model refers to a large model that uses quantization technology during the model operation process.
[0051] The types of input data (contextual data is comprehensively analyzed to determine the maximum actual value) and output data are determined based on the processing capabilities of the large model itself. For text and audio data, the maximum actual value can be determined based on the data length. For image data, the maximum actual value can be determined based on the value of the image in each channel.
[0052] S102: Determine the maximum representation capability of the quantized large model, and generate a corresponding scaling factor according to the maximum actual value and the maximum representation capability.
[0053] The maximum representation capability of each quantized large model can be determined by the highest accuracy of the data that can be processed in the large model. The main purpose of the scaling factor is to scale the input data so that the scaled input data is lower than the maximum representation capability.
[0054] Specifically, let the maximum actual value be V, let the maximum capability be Q, then first determine the difference between the maximum actual value V and the maximum capability Q. If the maximum actual value is higher than the maximum capability, it means that the input data needs to be scaled. At this time, the corresponding proportional coefficient is obtained by dividing the maximum capability and the maximum actual value. The proportional coefficient is represented as A, that is, A = Q / V. The proportional coefficient A can represent the proportion of the maximum capability Q to the maximum actual value V, that is, the highest proportion of the input data that the large model can currently represent.
[0055] Here, the proportional coefficient can be directly used as a scaling factor. After all input data are multiplied by the proportional coefficient A, all input data can be lower than or equal to the maximum capacity representation Q. Of course, a certain margin can also be reserved to cope with the possible numerical growth in vector operations.
[0056] At this time, the correction coefficient is determined according to the model depth of the quantized large model. Here, the correction coefficient is represented as P, and the proportional coefficient A is corrected by the correction coefficient P to reduce the proportional coefficient and obtain the corresponding scaling factor. For example, the scaling factor is represented as S, then S=P*A, where P is a positive number less than 1.
[0057] First, it is stipulated that the lower the value of the scaling factor S, the higher the scaling ability (of course, if the scaling factor is used as the denominator when scaling the input data, the higher the value of the scaling factor, the higher the scaling ability, so at this time it is only explained by using the scaling factor as the multiplication coefficient when scaling).
[0058] When the scaling factor S is equal to the proportional coefficient A, the maximum input data is multiplied by the scaling factor S, and after the input data is scaled, its maximum value is equal to the maximum capacity representation Q. If the scaling factor S is less than the proportional coefficient A after correction by the correction factor P, and its scaling capacity is increased by reducing the value of the scaling factor S, then the maximum input data is multiplied by the scaling factor S, and after the input data is scaled, its maximum value is lower than the maximum capacity representation Q, and a certain margin is also left.
[0059] Furthermore, when determining the correction coefficient, the model depth may be first obtained according to the number of intermediate layers of the quantized large model, wherein a higher number of intermediate layers often indicates a higher model depth.
[0060] If the model depth is lower than the preset depth, it means that the model depth of the current large model is low, and the possibility of numerical growth in its vector operation is low and the degree of numerical growth is low. Therefore, the correction coefficient is determined to be the preset first fixed value. For example, the correction coefficient is set to 0.9. At this time, the scaling factor S=0.9*A, which reduces the value of the scaling factor and increases its scaling ability.
[0061] If the model depth is higher than the preset depth, it means that in the current large model, the possibility of numerical growth in vector operations is high and the degree of numerical growth is high, so a higher margin is needed. At this time, according to the degree to which the model depth is higher than the preset depth, the preset first fixed value is reduced accordingly to obtain a correction coefficient. For example, for every 10% increase in the model depth above the preset depth, the first fixed value of 0.9 is reduced by 0.05. Assuming that the model depth is 25% higher than the preset depth, the degree to which it is higher than the preset depth reaches 20%, but less than 30%, so the correction coefficient is reduced by 0.1 on the basis of the first fixed value of 0.9 to 0.8. At this time, the scaling factor S = 0.8*A, which further increases its scaling ability compared to the first fixed value of 0.9, leaving more margin for large models.
[0062] Of course, if the maximum actual value is lower than the maximum representation capacity, it means that the input data of the current large model is lower than the maximum representation capacity of the large model. At this time, the preset second fixed value can be used as a scaling factor, wherein the second fixed value can be set to a value less than 1 and higher, for example, set to 0.9, or even the second fixed value can be directly set to 1.
[0063] S103: Scale the input data by the scaling factor so that the scaled input data is lower than the maximum capability representation.
[0064] After scaling the input data, it is equivalent to compressing the input data to a certain extent, so that the scaled input data is lower than the maximum capacity representation. Figure 2 As shown, a numerical mapping table can be established to restore the output data according to the numerical mapping table. The numerical mapping table includes the original numerical value corresponding to the input data and the corresponding scaled numerical value after scaling, and records the corresponding scaling factor.
[0065] Of course, for input data whose maximum actual value is higher than the maximum representation capability, scaling can be performed directly. However, for input data whose maximum actual value is lower than the maximum representation capability, scaling is not necessary. However, considering the data growth that may be caused by vector operations during the model calculation process, data scaling can still be performed at this time.
[0066] Specifically, all input data can be directly scaled. However, since the current maximum actual value is not higher than the maximum representation capacity, directly scaling all input data may easily lead to data distortion, which is not conducive to model training.
[0067] Based on this, key data and non-key data in the input data are obtained; wherein the key data and non-key data are determined by data tags of the input data.
[0068] At this time, the non-critical data is scaled by the scaling factor, and the critical data is not scaled because it is more critical, and is directly input into the quantitative large model, thereby improving the training quality of the model.
[0069] Furthermore, the data labels for key data and non-key data can be manually labeled or automatically generated by relevant algorithms. For example, for each input data, the data source of the input data is determined, and the reliability of the data source is determined. For different data sources, the corresponding reliability is marked.
[0070] The business relevance of the input data is determined for the business type corresponding to the input data. For different businesses (e.g., image recognition business, text translation business, etc.), the corresponding business relevance is set for different business types (e.g., for image recognition business, whether the image contains a recognition target and the difference in the recognition target can be set as the corresponding business type; for text translation business, the corresponding business type is determined according to its corresponding language type).
[0071] The uniqueness of the input data is determined based on the data difference between the input data and other input data. For example, for text data, the uniqueness is determined based on whether there are uncommon words and the proportion of uncommon words. For audio data, the uniqueness can be determined based on the text recognition results corresponding to the audio data, or the difference between the features corresponding to the audio data (including time domain features, frequency domain features, spectrum features, etc.) and other data. For image data, the uniqueness can be determined based on its color features, texture features, etc.
[0072] Determine whether the input data is critical data based on reliability, business relevance, and uniqueness. For example, construct corresponding scoring rules for each judgment parameter in reliability, business relevance, and uniqueness, and then set corresponding weights for each judgment parameter in reliability, business relevance, and uniqueness. The final score is obtained through weighting, and then determine whether it is critical data based on the score. Generally speaking, the higher the reliability, business relevance, and uniqueness, the higher the score, and the greater the possibility of being rated as critical data.
[0073] S104: Input the scaled input data into the quantized large model, obtain the output data of the quantized large model, and numerically restore the output data according to the scaling factor.
[0074] During the reasoning process of the large model, the output values are monitored in real time to ensure that any intermediate results remain within the range that can be accurately expressed by the quantitative large model. After the reasoning is completed, the output data of the large model is reversed according to the pre-established numerical mapping table, and reversely scaled according to the scaling factor to restore it to the original numerical space to ensure the accuracy of the results.
[0075] Specifically, when scaling, when the input data is text data, the text data is converted into a word vector, and the first preset dimension of the word vector (for example, the semantic dimension, grammatical dimension, emotional dimension, etc. in the word vector) is scaled, thereby scaling its length in space. When restoring, it is only necessary to reversely restore the length of the dimension in space according to the scaling factor.
[0076] When the input data is audio data, for each sampling point of the audio data, the second preset dimension of the sampling point (eg, signal amplitude, signal duration, etc.) is scaled. Similarly, when restoring, it is only necessary to perform reverse restoration according to the scaling factor.
[0077] When the input data is image data, for each pixel value in the image data, the third preset dimension of the pixel value (for example, the RGB channel value of the pixel value is scaled) is scaled. Similarly, when restoring, it only needs to be reversely restored according to the scaling factor.
[0078] This solution has the following beneficial effects:
[0079] 1. Improve accuracy and stability: Through the numerical scaling strategy, the precision overflow problem of large quantitative models when processing large numerical contexts is effectively solved, ensuring the accuracy of inference results and the stability of model operation.
[0080] 2. Resource efficiency: Compared with the method of directly expanding the model representation range, it achieves effective processing of large values without significantly increasing the model calculation and storage burden, maintaining the lightweight advantage of the quantitative model.
[0081] 3. Versatility and flexibility: It is highly versatile and suitable for various types of large quantitative models and different application scenarios. At the same time, the selection of scaling factors provides flexibility for different accuracy requirements.
[0082] 4. Easy to implement and integrate: With clear pre-processing, inference monitoring, and post-processing steps, the solution is easy to implement and integrate with existing quantitative model frameworks and inference engines.
[0083] like Figure 3 As shown, the embodiment of the present application also provides a large model precision overflow processing device based on numerical scaling, including:
[0084] at least one processor; and,
[0085] a memory communicatively connected to the at least one processor; wherein,
[0086] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the large model precision overflow processing method based on numerical scaling as described in any of the above embodiments.
[0087] An embodiment of the present application also provides a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to be: the large model precision overflow processing method based on numerical scaling as described in any of the above embodiments.
[0088] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.
[0089] The devices and media provided in the embodiments of the present application correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.
[0090] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0091] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0092] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0093] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0094] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0095] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.
[0096] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media 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 memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.
[0097] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
[0098] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.
Claims
1. A large model precision overflow processing method based on numerical scaling, characterized in that: include: Obtain input data of the current quantitative large model, and determine the maximum actual value thereof based on the input data; Determining the maximum representation capability of the quantized large model, and generating a corresponding scaling factor according to the maximum actual value and the maximum representation capability; Scaling the input data by the scaling factor so that the scaled input data is lower than the maximum capability representation; The scaled input data is input into the quantized large model, and the output data of the quantized large model is obtained, and the output data is numerically restored according to the scaling factor.
2. The method according to claim 1, characterized in that After scaling the input data by the scaling factor, the method further comprises: A numerical mapping table is established to facilitate numerical restoration of the output data according to the numerical mapping table; the numerical mapping table contains original numerical values corresponding to the input data and the corresponding scaled numerical values after scaling.
3. The method according to claim 1, characterized in that Generating a corresponding scaling factor according to the maximum actual value and the maximum representation capability includes: If the maximum actual value is higher than the maximum representation capability, the corresponding proportionality coefficient is obtained by dividing the maximum representation capability and the maximum actual value; Determining a correction coefficient according to the model depth of the quantitative large model; The proportional coefficient is corrected by the correction coefficient so that the proportional coefficient is reduced to obtain a corresponding scaling factor; wherein the lower the value of the scaling factor, the higher the scaling capability.
4. The method according to claim 3, characterized in that According to the model depth of the quantitative large model, a correction coefficient is determined, specifically including: Obtaining a model depth according to the number of intermediate layers of the quantized large model; If the model depth is lower than a preset depth, determining the correction coefficient to be a preset first fixed value; If the model depth is higher than a preset depth, the preset first fixed value is correspondingly reduced according to the degree to which the model depth is higher than the preset depth to obtain a correction coefficient.
5. The method according to claim 1, characterized in that Generating a corresponding scaling factor according to the maximum actual value and the maximum representation capability includes: If the maximum actual value is lower than the maximum representation capability, a preset second fixed value is used as the scaling factor.
6. The method according to claim 5, characterized in that Scaling the input data by the scaling factor specifically includes: Acquire key data and non-key data in the input data; wherein the key data and non-key data are determined by data tags of the input data; The non-critical data is scaled by the scaling factor, and the critical data is not scaled and directly input into the quantitative large model.
7. The method according to claim 6, characterized in that The method further comprises: For each input data, determine the data source of the input data and determine the reliability of the data source; and for the business type corresponding to the input data, determine the business relevance of the input data; and determine the uniqueness of the input data based on the data difference between the input data and other input data; Determine whether the input data is critical data based on the reliability, the business relevance, and the uniqueness.
8. The method according to claim 1, characterized in that: Scaling the input data by the scaling factor specifically includes: When the input data is text data, converting the text data into a word vector and scaling a first preset dimension of the word vector; When the input data is audio data, for each sampling point of the audio data, scaling the second preset dimension of the sampling point; When the input data is image data, for each pixel value in the image data, the third preset dimension of the pixel value is scaled.
9. A large model precision overflow processing device based on numerical scaling, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the large model precision overflow processing method based on numerical scaling as described in any one of claims 1 to 8.
10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are set to: the large model precision overflow processing method based on numerical scaling as described in any one of claims 1 to 8.