WO2025185502A9 - Data processing method and apparatus - Google Patents

Abstract

A data processing method, applied to the field of artificial intelligence, and comprising: acquiring first data, wherein the first data is obtained by means of a machine learning model on the basis of first input data; performing non-uniform quantization processing on the first data to obtain first compressed data, and storing the first compressed data in a memory; and reading the first compressed data from the memory, and performing inverse quantization processing corresponding to the non-uniform quantization processing on the first compressed data to obtain second data, wherein the second data and second input data are used for being input into the machine learning model, and the second input data is data input into the machine learning model after the first input data. According to the present application, non-uniform quantization is performed on data on the basis of non-uniform distribution characteristics of the data, so that the distribution characteristics of the data can be better met, and a quantization result having a smaller overall error or average error can be obtained, thereby improving the processing precision of models.

Description

A data processing method and apparatus

This application claims priority to Chinese Patent Application No. 202410245713.8, filed on March 4, 2024, entitled “A Data Processing Method and Apparatus”, the entire contents of which are incorporated herein by reference.

Technical Field

This application relates to the field of artificial intelligence, and more particularly to a data processing method and apparatus thereof.

Background Technology

Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a branch of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

When processing input data using a machine learning model, the machine learning model may include an attention layer. This attention layer performs attention calculations on the input data (e.g., data represented by tokens; for convenience, we will refer to this as tokens below). Furthermore, when performing attention calculations on the tokens, the attention layer obtains intermediate results that need to be reused in subsequent attention calculations on the tokens. For example, these intermediate results can be key-value (KV) data, i.e., a KV cache.

In this process, when processing a new token, the reusable intermediate results can be stored in memory so that when attention calculations are needed for other tokens later, the intermediate results can be read from memory and used as the basis for attention calculations for subsequent tokens.

However, as the size of the input data continues to increase, the amount of reusable intermediate results that need to be stored grows rapidly as inference progresses, resulting in a large storage requirement. Furthermore, excessively large intermediate results can also make the inference process extremely slow, thus necessitating the compression of reusable intermediate results.

Summary of the Invention

In a first aspect, this application provides a data processing method, the method comprising: acquiring first data, the first data being obtained by a machine learning model based on first input data; performing non-uniform quantization processing on the first data to obtain first compressed data, and storing the first compressed data in a memory; reading the first compressed data from the memory, and performing inverse quantization processing on the first compressed data corresponding to the non-uniform quantization processing to obtain second data, the second data and second input data being used as inputs into the machine learning model, the second input data being data input into the machine learning model after the first input data.

It should be understood that after performing non-uniform quantization on the first data, other processing (such as entropy coding) can be performed to obtain the first compressed data, and this application embodiment is not limited to this. Similarly, after performing the inverse transform processing corresponding to the nonlinear transform and the inverse quantization processing corresponding to the uniform quantization processing on the first compressed data, other processing can be performed to obtain the second data, and this application embodiment is not limited to this.

The second input data can be data input into the machine learning model after the first input data. For example, the second input data can be data input into the machine learning model after and adjacent to the first input data, or data input into the machine learning model after the first input data but separated by multiple input data.

The first data can be obtained by linearly transforming the input token by the intermediate network layer (e.g., attention layer) of the machine learning model. For example, the first data can be K data and/or V data.

Current caching and compression methods all use uniform quantization. However, the data output by large language models typically exhibits a non-uniform distribution, similar to a Gaussian distribution. Uniform quantization cannot well adapt to this non-uniform distribution characteristic, leading to a larger overall quantization error and significantly impacting small numerical values (primarily because data units tend to cluster within a small interval; uniform quantization converts these clustered units into the same or a small number of quantized values, resulting in significant accuracy loss). In this embodiment, non-uniform quantization is applied based on the non-uniform distribution characteristics of the data, better reflecting its distribution features and yielding quantization results with smaller overall or average errors, thereby improving the model's processing accuracy.

Taking key-value (KV) data as an example, current KV caching and compression methods all use uniform quantization. However, the distribution of KV data output by large language models generally exhibits a non-uniform distribution similar to a Gaussian distribution. Uniform quantization cannot well adapt to this non-uniform distribution characteristic of KV data, leading to a larger overall quantization error and a significant impact on small numerical values (mainly because KV data units are concentrated in a small interval; if uniform quantization is used, these concentrated data units will be converted into the same or a small number of quantized values, resulting in a significant loss of accuracy). In this embodiment, based on the non-uniform distribution characteristic of KV data, non-uniform quantization is performed on the KV data, which better matches the distribution characteristics of KV data, resulting in a quantization result with smaller overall or average error, thereby improving the processing accuracy of the model.

Specifically, in this embodiment, finer-grained quantization can be performed on denser numerical intervals in the first data, that is, more quantized values and corresponding intervals can be inserted (since there are more quantized values, the numerical width of the intervals will be lower), so that the quantization can better conform to the distribution characteristics of KV data and obtain a quantization result with smaller overall error or average error. For example, the first data includes multiple data units, the first compressed data includes quantized values corresponding to each data unit, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the quantized value corresponding to the data interval in which each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, the non-uniform quantization processing of the first data to obtain the first compressed data includes: converting the data units of the first data into corresponding quantized values based on a preset mapping relationship (e.g., in the form of a table) to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and quantized values corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, the step of performing non-uniform quantization processing on the first data to obtain first compressed data includes: performing a non-linear transformation on the first data to obtain transformed first data; and performing uniform quantization processing on the transformed first data.

The function type can be a power function, a logarithmic function, a gamma mapping function, an A-law curve, a μ-law curve, or a PWL curve, etc.

About: Power functions

Definition: Represented as x raised to the power of n

Mathematical expression: y = x^ ⁿ , where n is a real number

About: Logarithmic Function

Definition: The logarithmic function is the inverse function of the exponential function.

Mathematical expression: y = log _b (x) where b is the base of the logarithm.

About: Gamma Correction Function

Definition: Similar to a power function, it is typically used to map data onto normalized data.

Mathematical expression: y = x + ^γ , where x is the input pixel value and γ is the Gamma value, usually a real number greater than 0.

About: A-law Curve

Definition: The A-law curve is a non-linear coding curve, often used for audio signal compression.

Mathematical expression: y = ln(1 + A * |x|) / ln(1 + A) where x is the input signal and A is a constant.

About: μ-law curve

Definition: The μ-law curve is also a non-linear coding curve, often used for audio signal compression.

Mathematical expression: y＝sgn(x)*(ln(1+μ*|x|)/ln(1+μ)) where x is the input signal and μ is a constant.

About: PWL curve (Piecewise Linear Curve)

Definition: A PWL curve is a piecewise linear function composed of a series of line segments.

Mathematical expression:

Where ( _x1 , _y1 ), ( _x2 , _y2 )...( _xn , _yn ) are control points on the curve, and _m1 , _m2 ,... _mn are the slopes between adjacent control points.

In one possible implementation, the method further includes: determining the type of nonlinear function or the parameter values included in the nonlinear transformation based on the quantization precision of the first data.

Using the above method, the non-uniform mapping relationship in non-uniform quantization can be adaptively determined according to the quantization accuracy, further reducing quantization error. This reduces the overall quantization error while maintaining the compression gains brought by quantization.

In one possible implementation, the method further includes: determining the type of nonlinear function or the parameter values included in the nonlinear transformation based on the position of the network layer containing the attention layer in the machine learning model.

In one possible implementation, the first data is obtained from the first input data through the target head in the attention layer of the machine learning model; the method further includes: determining the type of nonlinear function or the parameter values included in the nonlinear transformation based on the position of the target head in the attention layer.

In one possible implementation, the type of nonlinear function or the parameter values used when performing the nonlinear transformation can be determined based on the generation interval (i.e., which generation order is determined by the storage order in memory, which can be called the sequence number) between the first data and the latest data obtained by the machine learning model.

In this way, the non-uniform mapping relationship in non-uniform quantization can be adaptively determined based on data grouping information (e.g., the position of the network layer containing the attention layer in the machine learning model, the position of the target head in the attention layer, and the generation order, at least one of these), further reducing quantization error. This reduces the overall quantization error while maintaining the compression benefits of quantization.

In one possible implementation, the method further includes: determining the category of the nonlinear function or the parameter values included in the nonlinear transformation based on the data distribution of the first data; the data distribution is indicated by the distribution statistics of the first data or by an identifier, wherein different identifiers correspond to data distributions with different characteristics.

The non-uniform mapping relationship in non-uniform quantization can be adaptively determined based on data distribution information, further reducing quantization error. This allows for a reduction in overall quantization error while maintaining the compression gains brought by quantization.

Secondly, this application provides a data processing method, the method comprising:

Obtain first data, which is calculated by a machine learning model based on first input data;

The first data is transformed by a nonlinear function to obtain the transformed first data.

The transformed first data is subjected to uniform quantization to obtain first compressed data, and the first compressed data is stored in a memory. The first compressed data is read from the memory, and the first compressed data is subjected to inverse transformation processing corresponding to the nonlinear transformation and inverse quantization processing corresponding to the uniform quantization processing to obtain second data. The second data and the second input data are used to input into the machine learning model. The second input data is the data input into the machine learning model after the first input data.

In one possible implementation, the category of the nonlinear function or the numerical values of its included parameters are determined based on at least one of the following:

The quantization precision of the first data; or,

The position of the network that generated the first data within the machine learning model; or,

The order in which the first data was generated in the machine learning model; or,

The data distribution of the first data.

It should be understood that the above-mentioned method for determining the nonlinear function and the method of uniform quantization combined with the nonlinear function can be applied to the compression and decompression process of other data besides KV data. "Other data" can be, but is not limited to, data that needs to be reused in the inference process of machine learning models.

In one possible implementation, the first data includes multiple data units, the first compressed data includes a quantized value corresponding to each data unit, the data units of the first data include data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, and the quantized value corresponding to the data interval in which each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belonging to the first numerical range, the second numerical interval belonging to the second numerical range, and the numerical width of the first numerical interval being smaller than the numerical width of the second numerical interval.

In one possible implementation, the step of performing non-uniform quantization on the first data to obtain the first compressed data includes:

Based on a preset mapping relationship, the data units of the first data are converted into corresponding quantized values to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and the quantized value corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, the step of performing non-uniform quantization on the first data to obtain the first compressed data includes:

The first data is subjected to a nonlinear transformation to obtain the transformed first data.

The transformed first data is then subjected to uniform quantization.

Thirdly, this application provides a data processing apparatus, the apparatus comprising:

An acquisition module is used to acquire first data, which is obtained by a machine learning model based on first input data; and to read first compressed data from the memory.

The processing module is used to perform non-uniform quantization processing on the first data to obtain the first compressed data, store the first compressed data in a memory, and perform inverse quantization processing on the first compressed data corresponding to the non-uniform quantization processing to obtain the second data. The second data and the second input data are used to input into the machine learning model. The second input data is the data input into the machine learning model after the first input data.

The processing module can be further divided into finer-grained parts. For example, the processing module may include a compression module and a decompression module. The compression module can perform non-uniform quantization processing on the first data to obtain the first compressed data. The decompression module can perform inverse transformation processing corresponding to the nonlinear transformation and inverse quantization processing corresponding to the uniform quantization processing on the first compressed data to obtain the second data.

In one possible implementation, the first data includes multiple data units, the first compressed data includes a quantized value corresponding to each data unit, the data units of the first data include data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, and the quantized value corresponding to the data interval in which each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belonging to the first numerical range, the second numerical interval belonging to the second numerical range, and the numerical width of the first numerical interval being smaller than the numerical width of the second numerical interval.

In one possible implementation, the processing module is specifically used for:

Based on a preset mapping relationship, the data units of the first data are converted into corresponding quantized values to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and the quantized value corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, the processing module is specifically used for:

The first data is subjected to a nonlinear transformation to obtain the transformed first data.

The transformed first data is then subjected to uniform quantization.

In one possible implementation, the processing module is further configured to:

Based on the quantization precision of the first data, determine the type of nonlinear function or the parameter values included in it when performing the nonlinear transformation.

In one possible implementation, the processing module is further configured to:

Based on the position of the network layer containing the attention layer in the machine learning model, the type of nonlinear function or the parameter values included in the nonlinear transformation are determined.

In one possible implementation, the first data is obtained from the first input data through the target head in the attention layer of a machine learning model; the processing module is further configured to:

Based on the position of the target head in the attention layer, determine the type of nonlinear function or the parameter values included in the nonlinear transformation.

In one possible implementation, the processing module is further configured to:

Based on the generation interval between the first data and the latest data obtained by the machine learning model, the category of the nonlinear function or the parameter values included in the nonlinear transformation are determined.

In one possible implementation, the processing module is further configured to:

Based on the data distribution of the first data, the category of the nonlinear function or the parameter values included in the nonlinear transformation are determined; the data distribution is indicated by the distribution statistics of the first data or by an identifier, wherein different identifiers correspond to data distributions with different characteristics.

Fourthly, this application provides a data processing apparatus, the apparatus comprising:

An acquisition module is used to acquire first data, which is calculated based on first input data through a machine learning model, and to read the first compressed data from the memory;

The processing module is configured to perform a nonlinear transformation on the first data using a nonlinear function to obtain the transformed first data; perform uniform quantization processing on the transformed first data to obtain the first compressed data, store the first compressed data in a memory, and perform inverse transformation processing corresponding to the nonlinear transformation and inverse quantization processing corresponding to the uniform quantization processing on the first compressed data to obtain second data. The second data and the second input data are used to input into the machine learning model, and the second input data is the data input into the machine learning model after the first input data.

In one possible implementation, the category of the nonlinear function or the numerical values of its included parameters are determined based on at least one of the following:

The quantization precision of the first data; or,

The position of the network that generated the first data within the machine learning model; or,

The order in which the first data was generated in the machine learning model; or,

The data distribution of the first data.

In one possible implementation, the first data includes multiple data units, the first compressed data includes a quantized value corresponding to each data unit, the data units of the first data include data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, and the quantized value corresponding to the data interval in which each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belonging to the first numerical range, the second numerical interval belonging to the second numerical range, and the numerical width of the first numerical interval being smaller than the numerical width of the second numerical interval.

In one possible implementation, the processing module is specifically used for:

Based on a preset mapping relationship, the data units of the first data are converted into corresponding quantized values to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and the quantized value corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

Fifthly, embodiments of this application provide a data processing apparatus, which may include a memory, a processor, and a bus system, wherein the memory is used to store a program, and the processor is used to execute the program in the memory to perform the methods described in the first aspect above and any optional methods thereof, or the methods described in the second aspect above and any optional methods thereof.

In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the methods described in the first aspect and any optional methods thereof, or the methods described in the second aspect and any optional methods thereof.

In a seventh aspect, embodiments of this application provide a computer program that, when run on a computer, causes the computer to perform the methods described in the first aspect and any optional methods thereof, or the methods described in the second aspect and any optional methods thereof.

Eighthly, this application provides a chip system including a processor for supporting an execution data processing device in implementing the functions involved in the foregoing aspects, such as transmitting or processing data involved in the foregoing methods; or, information. In one possible design, the chip system further includes a memory for storing program instructions and data necessary for the execution device or training device. This chip system may be composed of chips or may include chips and other discrete devices.

Attached Figure Description

Figure 1 is a schematic diagram of a structural framework for artificial intelligence.

Figures 2 to 4 are schematic diagrams of the application system framework of the present invention;

Figure 5 is a flowchart illustrating a data processing method provided in an embodiment of this application;

Figures 6A and 6B are schematic diagrams of a network structure provided in an embodiment of this application;

Figures 6C and 6D illustrate the data distribution.

Figure 6E is a schematic diagram of a nonlinear function;

Figures 7A to 7D are schematic diagrams of a data processing method provided in an embodiment of this application;

Figures 7E and 7F are schematic diagrams of the effects provided by an embodiment of this application;

Figure 8 is a schematic diagram of a data processing device provided in an embodiment of this application;

Figure 9 is a schematic diagram of an execution device provided in an embodiment of this application;

Figure 10 is a schematic diagram of a training device provided in an embodiment of this application;

Figure 11 is a schematic diagram of a chip structure provided in an embodiment of this application.

Detailed Implementation

The embodiments of the present invention will now be described with reference to the accompanying drawings. The terminology used in the embodiments section is for illustrative purposes only and is not intended to limit the scope of the invention.

The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

The terms “substantially,” “about,” and similar terms used herein are used as approximations rather than as terms of degree, and are intended to take into account the inherent biases of measurements or calculations known to those skilled in the art. Furthermore, the use of “may” in describing embodiments of the invention refers to “one or more possible embodiments.” The terms “use,” “using,” and “used” used herein are to be considered synonymous with the terms “utilize,” “utilizing,” and “utilized,” respectively. Additionally, the term “exemplary” is intended to refer to an instance or illustration.

First, the overall workflow of an artificial intelligence system is described, as shown in Figure 1. Figure 1 is a structural diagram of the main framework of artificial intelligence. The framework is then elaborated on from two dimensions: the "Intelligent Information Chain" (horizontal axis) and the "IT Value Chain" (vertical axis). The "Intelligent Information Chain" reflects a series of processes from data acquisition to processing. For example, it could be the general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, and intelligent execution and output. In this process, data undergoes a condensation process of "data—information—knowledge—wisdom." The "IT Value Chain" reflects the value that artificial intelligence brings to the information technology industry, from the underlying infrastructure of human intelligence and information (provided and processed by technology) to the industrial ecosystem of the system.

(1) Infrastructure

Infrastructure provides computing power to support artificial intelligence systems, enabling communication with the external world and providing support through a basic platform. This communication occurs through sensors; computing power is provided by intelligent chips (hardware acceleration chips such as CPUs, NPUs, GPUs, ASICs, and FPGAs); and the basic platform includes distributed computing frameworks and related platform guarantees and support, which may include cloud storage and computing, interconnected networks, etc. For example, sensors communicate with the outside world to acquire data, and this data is provided to intelligent chips in the distributed computing system provided by the basic platform for computation.

(2) Data

The data at the next layer of infrastructure is used to represent the data sources in the field of artificial intelligence. The data involves graphics, images, voice, text, and IoT data from traditional devices, including business data from existing systems and sensor data such as force, displacement, liquid level, temperature, and humidity.

(3) Data processing

Data processing typically includes methods such as data training, machine learning, deep learning, search, reasoning, and decision-making.

Among them, machine learning and deep learning can perform intelligent information modeling, extraction, preprocessing, and training on data, including symbolization and formalization.

Reasoning refers to the process in which, in a computer or intelligent system, the machine thinks and solves problems by simulating human intelligent reasoning, based on reasoning control strategies and using formalized information. Typical functions include search and matching.

Decision-making refers to the process of making decisions based on intelligent information after reasoning, and it typically provides functions such as classification, sorting, and prediction.

(4) General ability

After the data processing mentioned above, the results of the data processing can be used to form some general capabilities, such as algorithms or a general system, for example, translation, text analysis, computer vision processing, speech recognition, image recognition, etc.

(5) Smart Products and Industry Applications

Intelligent products and industry applications refer to products and applications of artificial intelligence systems in various fields. They are the encapsulation of overall artificial intelligence solutions, productizing intelligent information decision-making and realizing practical applications. Their application areas mainly include: intelligent terminals, intelligent transportation, intelligent healthcare, autonomous driving, smart cities, etc.

The system architecture provided in the embodiments of this application will be described in detail below with reference to Figure 2.

Figure 2 is a schematic diagram of the system architecture provided in an embodiment of this application. As shown in Figure 2, the system architecture 500 includes an execution device 510, a training device 520, a database 530, a client device 540, a data storage system 550, and a data acquisition system 560.

The execution device 510 includes a calculation module 511, an I/O interface 512, a preprocessing module 513, and a preprocessing module 514. The calculation module 511 may include a target model/rule 501, while the preprocessing modules 513 and 514 are optional.

The data acquisition device 560 is used to collect training samples. After collecting the training samples, the data acquisition device 560 stores these training samples in the database 530.

The training device 520 can maintain training samples in the database 530 to obtain the target model/rule 501 from the neural network to be trained (e.g., the machine learning model in the embodiments of this application).

It should be understood that the training device 520 can perform a pre-training process on the neural network to be trained based on the training samples maintained in the database 530, or fine-tune the model based on the pre-training.

It should be noted that in practical applications, the training samples maintained in database 530 may not all come from the data acquisition device 560; they may also be received from other devices. Furthermore, it should be noted that training device 520 may not necessarily train the target model/rule 501 entirely based on the training samples maintained in database 530; it may also obtain training samples from the cloud or other sources for model training. The above description should not be construed as limiting the embodiments of this application.

The target model/rule 501 trained by the training device 520 can be applied to different systems or devices, such as the execution device 510 shown in Figure 2. The execution device 510 can be a terminal, such as a mobile terminal, tablet computer, laptop computer, augmented reality (AR)/virtual reality (VR) device, vehicle terminal, etc., or it can be a server, etc.

Specifically, the training device 520 can transfer the trained model to the execution device 510.

In Figure 2, the execution device 510 is configured with an input/output (I/O) interface 512 for data interaction with external devices. Users can input data to the I/O interface 512 through the client device 540.

Preprocessing modules 513 and 514 are used to preprocess the input data received from the I/O interface 512. It should be understood that preprocessing modules 513 and 514 may be absent, or only one preprocessing module may be used. When preprocessing modules 513 and 514 are absent, the calculation module 511 can be used directly to process the input data.

During the preprocessing of input data by the execution device 510, or during the calculation module 511 of the execution device 510 performing calculations and other related processes, the execution device 510 can call data, code, etc. in the data storage system 550 for corresponding processing, or store the data, instructions, etc. obtained from the corresponding processing into the data storage system 550.

Finally, the I/O interface 512 provides the processing result to the client device 540, thereby providing it to the user.

In the scenario shown in Figure 2, the user can manually provide input data, which can be done through the interface provided by I/O interface 512. Alternatively, the client device 540 can automatically send input data to I/O interface 512. If user authorization is required for the client device 540 to automatically send input data, the user can set the corresponding permissions in the client device 540. The user can view the output results of the execution device 510 on the client device 540, which can be presented in various forms such as display, sound, or animation. The client device 540 can also act as a data acquisition terminal, collecting the input data and output results of the input I/O interface 512 as shown in the figure, and storing them as new sample data in database 530. Alternatively, data can be collected directly from the I/O interface 512 without going through the client device 540, using the input data and output results of the input I/O interface 512 as shown in the figure, and storing them as new sample data in database 530.

It is worth noting that Figure 2 is merely a schematic diagram of a system architecture provided in an embodiment of this application. The positional relationships between the devices, components, modules, etc., shown in the figure do not constitute any limitation. For example, in Figure 2, the data storage system 550 is an external memory relative to the execution device 510. In other cases, the data storage system 550 can also be placed in the execution device 510. It should be understood that the aforementioned execution device 510 can be deployed in the client device 540.

From the inference side of the model:

In this embodiment, the computing module 511 of the execution device 510 can obtain the code stored in the data storage system 550 to implement the steps related to the model reasoning process in this embodiment.

In this embodiment of the application, the computing module 511 of the execution device 510 may include hardware circuits (such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), general-purpose processors, digital signal processors (DSPs), microprocessors or microcontrollers, etc.) or combinations of these hardware circuits. For example, the training device 520 may be a hardware system with instruction execution capabilities, such as a CPU or DSP, or a hardware system without instruction execution capabilities, such as an ASIC or FPGA, or a combination of the aforementioned hardware systems without instruction execution capabilities and hardware systems with instruction execution capabilities.

Specifically, the computing module 511 of the execution device 510 can be a hardware system with the function of executing instructions. The steps related to the model inference process provided in this application embodiment can be software code stored in the memory. The computing module 511 of the execution device 510 can obtain the software code from the memory and execute the obtained software code to implement the steps related to the model inference process provided in this application embodiment.

It should be understood that the computing module 511 of the execution device 510 can be a combination of a hardware system without the function of executing instructions and a hardware system with the function of executing instructions. Some steps related to the model reasoning process provided in the embodiments of this application can also be implemented by the hardware system in the computing module 511 of the execution device 510 without the function of executing instructions, which is not limited here.

From the training side of the model:

In this embodiment, the training device 520 can obtain the code stored in the memory (not shown in Figure 2, which can be integrated into the training device 520 or deployed separately from the training device 520) to implement the steps related to model training in this embodiment.

In this embodiment of the application, the training device 520 may include hardware circuits (such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), general-purpose processors, digital signal processors (DSPs), microprocessors or microcontrollers, etc.) or combinations of these hardware circuits. For example, the training device 520 may be a hardware system with instruction execution capabilities, such as a CPU or DSP, or a hardware system without instruction execution capabilities, such as an ASIC or FPGA, or a combination of the aforementioned hardware systems without instruction execution capabilities and hardware systems with instruction execution capabilities.

It should be understood that the training device 520 can be a combination of a hardware system without the function of executing instructions and a hardware system with the function of executing instructions. Some steps related to the training of the neutralization model provided in the embodiments of this application can also be implemented by the hardware system in the training device 520 without the function of executing instructions, which is not limited here.

In this embodiment, the forward propagation process of the model is involved, which can be executed by the execution device 510 or the training device 520 described in the above embodiments.

Furthermore, the execution device 510 or training device 520 can process the input data using a machine learning model. This machine learning model may include an attention layer, which performs attention calculations on the input tokens. During these attention calculations, the attention layer can obtain intermediate results that can be reused in subsequent attention calculations on the same tokens. For example, these intermediate results can be K-data or V-data, and the K-data and V-data stored in memory constitute a KV cache. In this process, when processing a new token, reusable intermediate results can be stored in memory so that they can be retrieved from memory and used as a basis for attention calculations on other tokens. However, as the size of the input data increases, the amount of reusable intermediate results that need to be stored grows rapidly as inference progresses, leading to a large storage requirement. Furthermore, excessively large intermediate results can severely slow down the inference process; therefore, compressing reusable intermediate results is particularly important.

In one implementation, the compression process can be performed by a compression module, which can be centrally deployed with the execution device 510 or training device 520, for example, belonging to the same chip or other granular computing units, or it can be deployed separately, for example, belonging to different chips. For example, the execution device 510 or training device 520 can be an AI chip, and the compression module can belong to the CPU.

For example, referring to Figures 3 and 4, which are schematic diagrams of the architecture of an embodiment of this application, the model running module can obtain intermediate results by running a machine learning model, the compression module can compress the intermediate results and write the compressed data into the memory, and the compression module can read the compressed data from the memory and perform decompression to obtain the decompression result, which is then transmitted to the model running module. In Figure 3, the compression module and the model running module are deployed separately on different chips, while in Figure 4, the compression module and the model running module are centrally deployed on the same chip.

Since the embodiments of this application involve a large number of neural network applications, for ease of understanding, the relevant terms and concepts such as neural networks involved in the embodiments of this application will be introduced below.

(1) Neural Network

A neural network can be composed of neural units, which can be defined as a computational unit that takes xs (i.e., input data) and an intercept of 1 as input. The output of this computational unit can be:

Where s = 1, 2, ..., n, where n is a natural number greater than 1, Ws is the weight of xs, and b is the bias of the neural unit. f is the activation function of the neural unit, used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into an output signal. The output signal of this activation function can be used as the input of the next convolutional layer, and the activation function can be the sigmoid function. A neural network is a network formed by connecting multiple of the above-mentioned individual neural units together, that is, the output of one neural unit can be the input of another neural unit. The input of each neural unit can be connected to the local receptive field of the previous layer to extract the features of the local receptive field, which can be a region composed of several neural units.

(2) A convolutional neural network (CNN) is a deep neural network with a convolutional structure. A CNN contains a feature extractor consisting of convolutional layers and subsampling layers, which can be viewed as a filter. A convolutional layer refers to the layer of neurons in a CNN that performs convolutional processing on the input signal. In a convolutional layer of a CNN, a neuron can be connected to only some of the neurons in its neighboring layers. A convolutional layer typically contains several feature planes, each composed of a series of rectangularly arranged neural units. Neural units on the same feature plane share weights, which are the convolutional kernel. Shared weights can be understood as the way features are extracted being independent of their location. The convolutional kernel can be formalized as a matrix of random size, and during the training process of the CNN, the kernel can learn appropriate weights. Furthermore, the direct benefit of shared weights is reducing the connections between layers in the CNN, while also reducing the risk of overfitting.

CNN is a very common type of neural network. The following is a detailed introduction to the structure of CNN. As mentioned in the basic concept introduction above, a convolutional neural network is a deep neural network with a convolutional structure. It is a deep learning architecture, which refers to learning at multiple levels of different abstraction using machine learning algorithms. As a deep learning architecture, CNN is a feed-forward artificial neural network, where each neuron can respond to the input image.

(3) Deep Neural Networks

Deep Neural Networks (DNNs), also known as multilayer neural networks, can be understood as neural networks with many hidden layers, though there's no specific metric for "many." DNNs can be categorized into three layers based on their position: input layers, hidden layers, and output layers. Generally, the first layer is the input layer, the last layer is the output layer, and the layers in between are hidden layers. All layers are fully connected, meaning that any neuron in the i-th layer is connected to any neuron in the (i+1)-th layer. Although DNNs appear complex, the operation of each layer is actually quite simple, resembling a linear relationship as follows: in, It is the input vector. It is the output vector. α is the offset vector, W is the weight matrix (also called coefficients), and α() is the activation function. Each layer is simply an adjustment of the input vector. The output vector is obtained through such a simple operation. Because DNNs have many layers, the coefficients W and the offset vector... The number of these parameters is therefore quite large. The definitions of these parameters in a DNN are as follows: Taking the coefficient W as an example: Assuming a three-layer DNN, the linear coefficient from the 4th neuron in the second layer to the 2nd neuron in the third layer is defined as... The superscript 3 represents the layer number where coefficient W resides, while the subscript corresponds to the output third layer index 2 and the input second layer index 4. In summary, the coefficients from the k-th neuron in layer L-1 to the j-th neuron in layer L are defined as follows: It's important to note that the input layer does not have a W parameter. In deep neural networks, more hidden layers allow the network to better represent complex real-world situations. Theoretically, the more parameters a model has, the higher its complexity and "capacity," meaning it can perform more complex learning tasks. Training a deep neural network is essentially the process of learning the weight matrix, with the ultimate goal of obtaining the weight matrix of all layers in the trained deep neural network (a weight matrix formed by the vectors W from many layers).

(4) Loss Function

In training a deep neural network, to ensure the output closely approximates the desired predicted value, we compare the network's prediction with the target value. Based on the difference, we update the weight vector of each layer (usually pre-configuring parameters before the initial update). For example, if the prediction is too high, the weight vector is adjusted to predict a lower value. This process continues until the deep neural network predicts the target value or a value very close to it. Therefore, we need to predefine "how to compare the difference between the predicted and target values," which is the loss function or objective function. These are crucial equations used to measure the difference between the predicted and target values. Taking the loss function as an example, a higher output value (loss) indicates a greater difference, and training a deep neural network becomes a process of minimizing this loss.

(5) Backpropagation algorithm

Backpropagation (BP) can be used during training to correct the parameters in the initial model, thereby reducing the model's error loss. Specifically, forward propagation of the input signal to the output generates error loss; this error loss information is then propagated back to update the parameters in the initial model, leading to convergence of the error loss. The backpropagation algorithm is an error-loss-driven backpropagation process aimed at obtaining optimal model parameters, such as the weight matrix.

(6) Large Language Model: A large language model is a natural language processing model trained on large-scale data, typically with billions or tens of billions of parameters. These models learn the general features of language by studying a large amount of text data during the pre-training stage, and can then be fine-tuned on downstream tasks to adapt to the needs of specific tasks.

(7) Transformer: The transformer is a deep learning model architecture originally used for sequence-to-sequence tasks, such as machine translation. It uses a self-attention mechanism to process input sequences and has achieved great success in the field of natural language processing. Most large language models, such as BERT, GPT, and T5, are based on the Transformer architecture.

(8) Key-Value Cache: A key-value cache is a cache structure that stores key-value pairs. In large language models, key-value caches are often used to store intermediate results or other useful information that the model is processing text in order to improve efficiency. By using a key-value cache, the model can avoid redundant calculations when processing text.

(9) Key-Value Cache Quantization: Key-value cache quantization refers to quantizing the values in the key-value cache to reduce storage space and computational overhead. In some large language models, to adapt the model to limited resources, the values in the key-value cache can be quantized to reduce the model's storage and computational costs.

(10) PPL (Perplexity): PPL is a metric used to evaluate the performance of a language model, representing the model's ability to predict a given text sequence. PPL is a positive real number, which can be understood as the average difficulty the model has in predicting the observed data sequence. The lower the PPL, the better the model performance.

(11) Non-uniform quantization: Non-uniform quantization is a quantization method in which the numerical range is divided into intervals of different sizes to better adapt to the data distribution. Unlike uniform quantization, non-uniform quantization can assign different numbers of numerical ranges to each interval according to the distribution of the data.

(12) Token: In natural language processing, a "token" is the basic unit for segmenting a text string. This can be a word, a character, or a fragment of a word. Large language models typically need to segment the input text into tokens and then convert these tokens into numerical representations (such as word vectors) that the model can understand.

(13) Sequence: In the context of the large language model, "sequence" refers to a sequence of elements with a certain order relationship. Multiple tokens make up a sequence.

(14) Incremental Inference: Incremental inference allows the model to process only newly added parts of the input, rather than reprocessing the entire sequence each time. This is achieved by maintaining contextual information in the model's internal state, allowing the model to respond quickly when it receives new input. Incremental inference is particularly useful in interactive applications, such as chatbots or real-time translation, as it can significantly reduce latency and computational resource usage.

This application provides a data processing method. The data processing method of this application embodiment will be described in detail below with reference to the accompanying drawings.

Referring to Figure 5, which is a flowchart of a data processing method provided in an embodiment of this application, the data processing method provided in this application may include steps 501 to 503, which will be described in detail below.

501. Obtain first data, which is obtained from the first input data through a machine learning model;

In the above process, each time the machine learning model processes the latest input data, it is necessary to obtain the intermediate results that have been generated in the past (such as K data or V data) and input them into the machine learning model to perform attention operations on the latest input data (that is, operations based on attention mechanisms). In this process, intermediate results will still be obtained (that is, data that still needs to be reused when processing subsequent input data). The newly generated intermediate results and the obtained intermediate results generated in the past can be concatenated and attention weights can be calculated.

For example, referring to Figure 6B, K cache and V cache represent historically generated intermediate results. After calculating the token of the latest input data (e.g., through linear transformations of the Wd, Wk, and Wv matrices), the Q, K, and V data of the latest input data can be obtained. Then, the K data and K cache of the latest input data can be concatenated and transposed, and the V data and V cache of the latest input data can be concatenated. Subsequent operations can be described in the relevant introduction to attention operations in existing technologies, which will not be repeated here. The K and V data of the latest input data can be stored in memory and retrieved from memory when the machine learning model performs operations on subsequent input data.

Among them, the machine learning model can be a language model.

The attention layer can perform self-attention calculations on the input data.

Taking a machine learning model based on an attention mechanism, such as the transformer model (see Figure 6A), as an example, the machine learning model could be a language model. The transformer is a model architecture based on an attention mechanism, with self-attention being its core component. Each transformer layer contains two main parts: a multi-head self-attention layer and a feedforward neural network (FNN). The self-attention structure allows the model to dynamically focus on information at different positions when processing sequential data. It consists of three main parts: query data, key data, and value data. For an input sequence, by calculating the linear transformations of Q, K, and V, and then performing a softmax operation, the attention distribution of each position to other positions is obtained. Finally, these distributions are weighted to obtain the output at the current position. The self-attention structure is used to process information at different positions in the input data. In self-attention, the representation of each position is obtained by a weighted average of all other positions in the sequence, with the weights determined by calculating the relationship between the query, key, and value at the current position. This allows the model to dynamically focus on different parts of the input sequence.

During inference, the self-attention structure generates a key-value (KV) pair for each position. These KV pairs are used in the attention mechanism to calculate the weights at different positions. The KV cache refers to the set of KV pairs generated at a certain time step. Storing historical KV pairs can avoid repeated calculations, thereby accelerating the model's inference.

In one possible implementation, the first data can be the data that the machine learning model described above needs to reuse during the computation of subsequent input data.

In one possible implementation, the first data can be K data or V data. That is, the first data can be K data, the first data can be V data, or the first data can be both K data and V data.

As the size of the input data increases, the amount of reusable intermediate results that need to be stored grows rapidly as inference progresses, leading to a significant storage requirement. Furthermore, excessively large intermediate results can severely slow down the inference process. Therefore, compressing reusable intermediate results becomes crucial. Before storing the intermediate results (i.e., the initial data) that need to be reused during attention calculation, compression processing is required, and the compressed data should be stored to reduce storage overhead. When using the compressed data, it needs to be read and decompressed to ensure the model's accuracy is maintained when calculating the attention distribution.

In one possible implementation, the first data may be obtained by compressing the intermediate result.

In one possible implementation, the first data is specifically obtained by compressing and decompressing the intermediate result.

Specifically, in one possible implementation, the first data can be obtained by a machine learning model based on the latest input data (first input data), or it can be the result obtained by compressing and decompressing the intermediate result obtained by a machine learning model based on the latest input data.

Specifically, in one possible implementation, the first data can be obtained through a machine learning model based on the latest input data, or it can be the result obtained by compressing and decompressing the intermediate results obtained through a machine learning model based on the latest input data.

Specifically, in one possible implementation, the first data may not be obtained from the latest input data (the first input data), but rather from compressed data retrieved from memory, or it may be obtained by decompressing compressed data retrieved from memory. It should be understood that this compression-decompression-recompression process only occurs when determining the non-uniform quantization correspondence (or the type and parameters of the nonlinear mapping function) based on the sequence number. That is, the K or V generated by the same token needs to undergo different quantization processes at different times. If the quantization parameters change at a certain time, decompression and requantization are required.

When the first data is obtained based on the latest input data, it needs to be compressed and stored in memory.

When the first data is obtained from memory instead of the latest input data, it may be necessary to compress the first data with different strengths (i.e., compression with different degrees of precision loss) and store it in memory.

502. Perform non-uniform quantization processing on the first data to obtain first compressed data, and store the first compressed data in the memory;

In this embodiment of the application, after obtaining the first data, the first data can be compressed to obtain the first compressed data, and the first compressed data can be stored in a memory, for example, a memory cache.

Current key-value (KV) caching compression methods all use uniform quantization. However, the distribution of KV data output by large language models generally exhibits a non-uniform distribution similar to a Gaussian distribution. As shown in Figures 6C and 6D, the key and value cache data output by a certain layer of the Llama-7B large language model exhibit a non-uniform distribution. Uniform quantization cannot well adapt to this non-uniform distribution characteristic of KV data, leading to a larger overall quantization error and a greater impact on small numerical values (mainly because KV data units are concentrated in a small interval; if uniform quantization is used, these concentrated data units will be converted into the same or a small number of quantized values, resulting in a significant loss of accuracy).

In this embodiment, based on the non-uniform distribution characteristics of KV data, non-uniform quantization of KV data can better conform to the distribution characteristics of KV data, resulting in quantization results with smaller overall or average errors, thereby improving the processing accuracy of the model.

Specifically, in this embodiment, finer-grained quantization can be performed on denser numerical intervals in the first data, that is, more quantized values and corresponding intervals can be inserted (since there are more quantized values, the numerical width of the intervals will be lower). For example, the first data includes multiple data units, the first compressed data includes quantized values corresponding to each data unit, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the quantized value corresponding to the data interval where each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, the first data can be nonlinearly transformed to obtain transformed first data, and then the transformed first data can be uniformly quantized. Specifically, the processing flow shown in Figure 7A can be referred to.

For example, the data is mapped and transformed according to a predetermined non-uniform mapping function (or non-linear transformation function). The non-uniform mapping function can be a power function, a logarithmic function, an A-law curve, a μ-law curve, or a piecewise linear curve, etc. For example, as shown in Figure 6E, which is a schematic diagram of a non-linear function; then the mapped data is uniformly quantized. The quantization coefficient of uniform quantization can be a preset value or can be adaptively determined by online statistical analysis of the mapped data.

Among them, the above-mentioned nonlinear transformation and uniform quantization steps can be combined. In this case, nonlinear mapping and uniform quantization can be combined into a table lookup step: according to the pre-set nonlinear transformation function and uniform quantization method, a correspondence table between floating-point value range and quantized value can be constructed. Quantization can be completed by matching the K data or V data with the table, and the quantized K data or V data can be obtained.

Specifically, in one possible implementation, the data units of the first data can be converted into corresponding quantized values based on a preset mapping relationship to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and quantized values corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, the type of nonlinear function or the parameter values included in the nonlinear transformation can be determined based on the quantization precision of the first data. Different data to be quantized, such as key-value cache data output from different layers of a large language model, key-value cache data at different head dimensions, and key-value cache data at different time points during model inference, may use different quantization precisions. The quantization mapping relationship can be adaptively determined based on the quantization precision (e.g., the number of quantization bits), and quantization can be performed according to the determined quantization mapping relationship.

For example, the K or V data to be cached, and the quantization precision (number of quantization bits) of the current K or V data can be obtained. The type and parameters of the nonlinear transformation function are determined based on the quantization precision. The numerical mapping relationship can be determined from the type and parameters of the nonlinear transformation function. The function type can be a power function, logarithmic function, gamma mapping function, A-law curve, μ-law curve, or PWL curve, etc. Nonlinear mapping is then performed on the KV cached data to be quantized. Uniform quantization is then performed on the transformed KV cached data, and the quantized K or V data is output.

Nonlinear mapping and uniform quantization can be combined into a single lookup step: given a defined nonlinear transformation function and uniform quantization method, a table mapping floating-point value ranges to quantized values can be constructed. Quantization can then be completed by matching the values of K-data or V-data with the table, yielding the quantized K-data or V-data.

The determination of the type and parameters of the nonlinear transformation function based on the quantization precision can be achieved through various methods: for example, it can be determined based on the correspondence between a preset quantization precision and the type and parameters of the nonlinear transformation function; alternatively, one quantization precision can correspond to multiple preset types and parameters of nonlinear transformation functions, and the type and parameters of the nonlinear transformation function corresponding to different quantization precisions can be determined by testing on a validation set. Specifically, the processing flow shown in Figure 7A can be referenced.

Using the above method, the non-uniform mapping relationship in non-uniform quantization can be adaptively determined according to the quantization accuracy, further reducing quantization error. This reduces the overall quantization error while maintaining the compression gains brought by quantization. Specifically, refer to the processing flow shown in Figure 7B.

In one possible implementation, the type of nonlinear function or the parameter values used when performing the nonlinear transformation can be determined based on the position of the network layer containing the attention layer in the machine learning model.

In one possible implementation, the first data is K data and V data obtained from the target head in the attention layer of the machine learning model based on the first input data; the type of nonlinear function or the parameter values included in the nonlinear transformation can be determined based on the position of the target head in the attention layer.

In one possible implementation, the type of nonlinear function or the parameter values used when performing the nonlinear transformation can be determined based on the generation interval (i.e., which generation order is determined by the storage order in memory, which can be called the sequence number) between the first data and the latest data obtained by the machine learning model.

Besides changing with quantization precision, the quantization mapping relationship also considers the grouping information of the data. For example, different quantization mapping relationships can be used for KV cached data output from different layers; different quantization mapping relationships can also be used for KV cached data with different head dimensions or sequence dimensions. Specifically, the required K or V data to be cached, the quantization precision (number of quantization bits) of the current K or V data, and the grouping information are obtained. The grouping information can be the sequence number of the layer outputting the current K or V data, the head dimension sequence number of the current K or V data, the sequence number of the current K or V data, the group sequence number of the current K or V data, or the statistical information of the validation set in the current group.

The type and parameters of the nonlinear transformation function are determined based on the quantization precision and data grouping information. The numerical mapping relationship can be determined from the type and parameters of the nonlinear transformation function. The function type can be a power function, logarithmic function, gamma mapping function, A-law curve, μ-law curve, or PWL curve, etc. Nonlinear mapping is then performed on the KV buffer data to be quantized. Uniform quantization is then performed on the transformed KV buffer data, outputting the quantized K or V data. Nonlinear mapping and uniform quantization can be combined into a single lookup step: based on the determined nonlinear transformation function and uniform quantization method, a table mapping floating-point value ranges to quantized values can be constructed. Quantization can then be completed by matching the K or V data values to the table, yielding the quantized K or V data. Determining the type and parameters of the nonlinear transformation function based on the quantization precision and data grouping information can be achieved through various methods: for example, it can be determined based on a preset correspondence between the quantization precision, grouping information, and the type and parameters of the nonlinear transformation function; alternatively, one quantization precision plus grouping information can correspond to multiple preset types and parameters of nonlinear transformation functions, and the type and parameters of the nonlinear transformation function corresponding to different quantization precisions and grouping information can be determined by testing on a validation set.

In this way, the non-uniform mapping relationship in non-uniform quantization can be adaptively determined based on data grouping information (e.g., the position of the network layer containing the attention layer in the machine learning model, the position of the target head in the attention layer, and the generation order, at least one of these), further reducing quantization error. This reduces the overall quantization error while maintaining the compression benefits of quantization. Specifically, the processing flow shown in Figure 7C can be referenced.

In one possible implementation, the type of nonlinear function or the parameter values included in the nonlinear transformation can be determined based on the data distribution of the first data; the data distribution is indicated by the distribution statistics of the first data or by an identifier, wherein different identifiers correspond to data distributions with different characteristics. For example, the quantization mapping relationship changes not only with the quantization precision but also considers the distribution information of the current data to be quantized. For example, the quantization mapping relationship can be determined based on the quantization precision and the dispersion of the current quantized data.

Specifically, the process involves obtaining the K or V data to be cached, the quantization precision (number of quantization bits) of the current K or V data, and the distribution information of the current K or V data. The data distribution information consists of statistical values from the current K or V data, such as mean, variance, maximum, minimum, and range. Based on the quantization precision and data distribution information, the type and parameters of the nonlinear transformation function are determined. The type and parameters of the nonlinear transformation function determine the numerical mapping relationship. The function type can be a power function, logarithmic function, gamma mapping function, A-law curve, μ-law curve, or PWL curve, etc. Nonlinear mapping is then performed on the KV cache data to be quantized. Uniform quantization is then performed on the transformed KV cache data, outputting the quantized K or V data. The data distribution information is saved for use in the dequantization stage. For details, refer to the processing flow shown in Figure 7D.

Nonlinear mapping and uniform quantization can be combined into a single lookup table step: given a defined nonlinear transformation function and uniform quantization method, a table mapping floating-point value ranges to quantized values can be constructed. Quantization can then be completed by matching the K or V data values to this table, yielding the quantized K or V data. The type and parameters of the nonlinear transformation function can be determined based on quantization precision and data distribution information, which can be achieved through various methods. For example, testing on a validation set can obtain calibrated distribution information. During the quantization and dequantization phases, the difference between the current data distribution information and the calibrated distribution information, along with the quantization precision, can then be used to determine the type and parameters of the nonlinear transformation function.

The non-uniform mapping relationship in non-uniform quantization can be adaptively determined based on data distribution information, further reducing quantization error. It can significantly reduce overall quantization error while maintaining the compression gains brought by quantization.

503. Read the first compressed data from the memory and perform inverse quantization processing on the first compressed data corresponding to the non-uniform quantization processing to obtain second data. The second data and the second input data are used to input into the machine learning model. The second input data is the data input into the machine learning model after the first input data.

Depending on how the nonlinear function is determined, dequantization can be performed in, but is not limited to, the following ways.

When KV Cache data is needed, the quantized KV Cache data needs to be dequantized. The specific steps of the dequantization stage are illustrated as follows: Obtain the quantized K data or V data; perform uniform dequantization on the quantized KV cache data using the same uniform quantization method as in the quantization stage; perform inverse transformation on the dequantized data using the inverse nonlinear transformation function corresponding to the nonlinear transformation in the quantization stage; obtain the decompressed KV Cache data for inference.

Among them, uniform dequantization and nonlinear inverse mapping can be combined into a table lookup step: based on the uniform dequantization method and the pre-set nonlinear inverse transformation function, a correspondence table between quantized values and floating-point values can be constructed. Dequantization can be completed by matching the values of the quantized K data or V data with the table, and the dequantized K data or V data can be obtained.

When KV Cache data is needed, the quantized KV Cache data needs to be dequantized. The specific steps of the dequantization stage are as follows: Obtain the quantized K data or V data, obtain its quantization precision, and perform uniform dequantization on the quantized KV cache data using the same uniform quantization method as in the quantization stage; determine the type and parameters of the nonlinear inverse transform function based on the quantization precision, and the function is the inverse function corresponding to the function used in the quantization stage; perform inverse transform on the dequantized data using the nonlinear inverse transform function; obtain the decompressed KV Cache data for inference.

Nonlinear inverse mapping and uniform dequantization can be combined into a single table lookup step: based on the uniform dequantization method and a determined nonlinear inverse transformation function, a table corresponding to quantized values and floating-point values can be constructed. Dequantization can then be completed by matching the values of the quantized K-data or V-data with the table, yielding the dequantized K-data or V-data.

When KV Cache data is needed, the quantized KV Cache data needs to be dequantized. The specific steps of the dequantization stage are as follows: Obtain the quantized K data or V data, obtain the quantization precision (number of quantized bits) of the current KV quantized data, and the grouping information. The grouping information can be the layer number, head dimension number, sequence number, or group number of the current K data or V data; perform uniform dequantization on the quantized KV cache data using the same uniform quantization method as in the quantization stage; determine the type and parameters of the nonlinear inverse transform function based on the quantization precision and data grouping information obtained in step 1, where the function is the inverse function corresponding to the function used in the quantization stage; perform inverse transform on the dequantized data using the nonlinear inverse transform function; obtain the decompressed KV Cache data for inference.

Nonlinear inverse mapping and uniform dequantization can be combined into a single table lookup step: based on the uniform dequantization method and the nonlinear inverse transformation function determined in step 3, a correspondence table between quantized values and floating-point values can be constructed. Dequantization can be completed by matching the values of the quantized K data or V data with the table, thus obtaining the dequantized K data or V data.

When KV Cache data is needed, it must be dequantized. The specific steps of the dequantization stage are as follows: Obtain the quantized K or V data, obtain the quantization precision (number of quantized bits) of the current KV quantized data, and the distribution information of the current K or V data. The data distribution information consists of statistical values from the current K or V data, such as mean, variance, maximum value, minimum value, and value range; perform uniform dequantization on the quantized KV cache data using the same uniform quantization method as in the quantization stage; determine the type and parameters of the nonlinear inverse transform function based on the obtained quantization precision and distribution information, where the function is the inverse function corresponding to the function used in the quantization stage; perform an inverse transform on the dequantized data using the nonlinear inverse transform function; and obtain the decompressed KV Cache data for inference.

Nonlinear inverse mapping and uniform dequantization can be combined into a single table lookup step: based on the uniform dequantization method and the determined nonlinear inverse transformation function, a correspondence table between quantized values and floating-point values can be constructed. Dequantization can then be completed by matching the values of the quantized K-data or V-data with the table, yielding the dequantized K-data or V-data.

The beneficial effects of the embodiments of this application will be described below with reference to experiments:

The wikitext2 validation set was tested on the Llama-7B model. The KV cache data generated during inference was quantized using the non-uniform quantization method proposed in this invention. When historical cached K or V data was needed during inference, it was dequantized. The compression ratio and perplexity (PPL) curves are shown in Figure 7E when the sequence length is 2048. As can be seen from the figure, compared with uniform quantization, non-uniform quantization achieves better inference performance while maintaining the same compression ratio (lower PPL indicates better inference performance).

The wikitext2 validation set was tested on the Llama-7B model. The KV cache data generated during inference was quantized using the method proposed in this invention, which "determines the quantization mapping relationship based on quantization precision." Historical cached K or V data was dequantized when needed during inference. The compression ratio and test PPL curves are shown in Figure 7F when the sequence length is 2048. As can be seen from the figure, compared to using preset non-uniform quantization, the quantization method that determines the quantization mapping relationship based on quantization precision achieves better inference performance while maintaining the same compression ratio (lower PPL indicates better inference performance).

It should be understood that the above-mentioned method for determining the nonlinear function and the method of uniform quantization combined with the nonlinear function can be applied to the compression and decompression process of other data besides KV data. "Other data" can be, but is not limited to, data that needs to be reused in the inference process of machine learning models.

Specifically, this application provides a data processing method, the method comprising: acquiring first data, the first data being calculated by a machine learning model based on first input data; performing a nonlinear transformation on the first data using a nonlinear function to obtain transformed first data; performing uniform quantization processing on the transformed first data to obtain first compressed data, and storing the first compressed data in a memory; reading the first compressed data from the memory, and performing an inverse transformation processing corresponding to the nonlinear transformation and an inverse quantization processing corresponding to the uniform quantization processing on the first compressed data to obtain second data, the second data and the second input data being used as inputs into the machine learning model, the second input data being data input into the machine learning model after the first input data; wherein, the category of the nonlinear function or the parameter values included therein are determined based on at least one of the following: the quantization precision of the first data; or, the position of the network that generated the first data in the machine learning model; or, the generation order of the first data in the machine learning model; or, the data distribution of the first data.

In one possible implementation, the first data includes multiple data units, the first compressed data includes a quantized value corresponding to each data unit, the data units of the first data include data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, and the quantized value corresponding to the data interval in which each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belonging to the first numerical range, the second numerical interval belonging to the second numerical range, and the numerical width of the first numerical interval being smaller than the numerical width of the second numerical interval.

In one possible implementation, the data units of the first data can be converted into corresponding quantized values based on a preset mapping relationship to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and quantized values corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, when performing non-uniform quantization on the first data, the first data can be non-linearly transformed to obtain the transformed first data; then, the transformed first data can be subjected to uniform quantization.

Referring to Figure 8, which is a schematic diagram of the structure of a data processing apparatus provided in an embodiment of this application, as shown in Figure 8, the data processing apparatus 800 provided in this embodiment of the application includes:

The acquisition module 801 is used to acquire first data, which is obtained by a machine learning model based on first input data, and reads first compressed data from the memory;

For a detailed description of the acquisition module 801, please refer to the description of step 501 in the above embodiment, which will not be repeated here.

The processing module 802 is used to perform non-uniform quantization processing on the first data to obtain the first compressed data, store the first compressed data in a memory, and perform inverse quantization processing on the first compressed data corresponding to the non-uniform quantization processing to obtain the second data. The second data and the second input data are used to input into the machine learning model. The second input data is the data input into the machine learning model after the first input data.

For a detailed description of the acquisition module 801, please refer to the descriptions of steps 502 and 503 in the above embodiments, which will not be repeated here.

In one possible implementation, the first data includes multiple data units, the first compressed data includes a quantized value corresponding to each data unit, the data units of the first data include data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, and the quantized value corresponding to the data interval in which each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belonging to the first numerical range, the second numerical interval belonging to the second numerical range, and the numerical width of the first numerical interval being smaller than the numerical width of the second numerical interval.

In one possible implementation, the processing module 802 is specifically used for:

Based on a preset mapping relationship, the data units of the first data are converted into corresponding quantized values to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and the quantized value corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

In one possible implementation, the processing module 802 is specifically used for:

The first data is subjected to a nonlinear transformation to obtain the transformed first data.

The transformed first data is then subjected to uniform quantization.

In one possible implementation, the processing module 802 is further configured to:

Based on the quantization precision of the first data, determine the type of nonlinear function or the parameter values included in it when performing the nonlinear transformation.

In one possible implementation, the processing module 802 is further configured to:

Based on the position of the network layer containing the attention layer in the machine learning model, the type of nonlinear function or the parameter values included in the nonlinear transformation are determined.

In one possible implementation, the first data is obtained from the first input data through the target head in the attention layer of a machine learning model; the processing module 802 is further configured to:

Based on the position of the target head in the attention layer, determine the type of nonlinear function or the parameter values included in the nonlinear transformation.

In one possible implementation, the processing module 802 is further configured to:

Based on the generation interval between the first data and the latest data obtained by the machine learning model, the category of the nonlinear function or the parameter values included in the nonlinear transformation are determined.

In one possible implementation, the processing module 802 is further configured to:

Based on the data distribution of the first data, the category of the nonlinear function or the parameter values included in the nonlinear transformation are determined; the data distribution is indicated by the distribution statistics of the first data or by an identifier, wherein different identifiers correspond to data distributions with different characteristics.

This application embodiment also provides a data processing apparatus, the apparatus comprising:

An acquisition module is used to acquire first data, which is calculated based on first input data through a machine learning model, and to read the first compressed data from the memory;

The processing module is configured to perform a nonlinear transformation on the first data using a nonlinear function to obtain transformed first data; perform uniform quantization on the transformed first data to obtain first compressed data, store the first compressed data in a memory, and perform inverse transformation processing corresponding to the nonlinear transformation and inverse quantization processing corresponding to the uniform quantization on the first compressed data to obtain second data. The second data and the second input data are used to input into the machine learning model, and the second input data is the data input into the machine learning model after the first input data.

In one possible implementation, the category of the nonlinear function or the numerical values of its included parameters are determined based on at least one of the following:

The quantization precision of the first data; or,

The position of the network that generated the first data within the machine learning model; or,

The order in which the first data was generated in the machine learning model; or,

The data distribution of the first data.

In one possible implementation, the first data includes multiple data units, the first compressed data includes a quantized value corresponding to each data unit, the data units of the first data include data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, and the quantized value corresponding to the data interval in which each data unit is located is used as the quantized value corresponding to the data unit, wherein the numerical interval includes a first numerical interval and a second numerical interval, the first numerical interval belonging to the first numerical range, the second numerical interval belonging to the second numerical range, and the numerical width of the first numerical interval being smaller than the numerical width of the second numerical interval.

In one possible implementation, the processing module is specifically used for:

Based on a preset mapping relationship, the data units of the first data are converted into corresponding quantized values to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and the quantized value corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval.

The following describes a terminal device provided in an embodiment of this application. Please refer to Figure 9, which is a structural schematic diagram of a terminal device provided in an embodiment of this application. The terminal device 900 can specifically be a virtual reality (VR) device, a mobile phone, a tablet, a laptop computer, a smart wearable device, etc., and is not limited here. Specifically, the terminal device 900 includes: a receiver 901, a transmitter 902, a processor 903, and a memory 904 (the number of processors 903 in the terminal device 900 can be one or more; Figure 9 shows one processor as an example). The processor 903 may include an application processor 9031 and a communication processor 9032. In some embodiments of this application, the receiver 901, transmitter 902, processor 903, and memory 904 can be connected via a bus or other means.

Memory 904 may include read-only memory and random access memory, and provides instructions and data to processor 903. A portion of memory 904 may also include non-volatile random access memory (NVRAM). Memory 904 stores processor and operation instructions, executable modules, or data structures, or subsets thereof, or extended sets thereof, wherein the operation instructions may include various operation instructions for implementing various operations.

Processor 903 controls the operation of the execution device. In specific applications, the various components of the execution device are coupled together through a bus system, which may include not only the data bus, but also power buses, control buses, and status signal buses. However, for clarity, all buses in the diagram are referred to as the bus system.

The methods disclosed in the embodiments of this application can be applied to or implemented by the processor 903. The processor 903 can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 903 or by instructions in software form. The processor 903 can be a general-purpose processor, a digital signal processor (DSP), a microprocessor, or a microcontroller, and may further include an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The processor 903 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 module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 904. The processor 903 reads the information from memory 904 and, in conjunction with its hardware, completes the steps involved in the model training or model inference process described above.

Receiver 901 can be used to receive input digital or character information, and to generate signal inputs related to the settings and function control of the execution device. Transmitter 902 can be used to output digital or character information through the first interface; transmitter 902 can also be used to send instructions to the disk group through the first interface to modify the data in the disk group; transmitter 902 may also include a display device such as a display screen.

This application embodiment also provides a server. Referring to Figure 10, which is a schematic diagram of a server structure provided in this application embodiment, the server 1000 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 1010 (e.g., one or more processors) and memory 1032, and one or more storage media 1030 (e.g., one or more mass storage devices) for storing application programs 1042 or data 1044. The memory 1032 and storage media 1030 can be temporary or persistent storage. The program stored in the storage media 1030 may include one or more modules (not shown in the figure), each module may include a series of instruction operations on the server. Furthermore, the CPU 1010 may be configured to communicate with the storage media 1030 and execute the series of instruction operations in the storage media 1030 on the server 1000.

Server 1000 may also include one or more power supplies 1026, one or more wired or wireless network interfaces 1050, one or more input/output interfaces 1058; or, one or more operating systems 1041, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

In this embodiment, the central processing unit 1010 is used to perform actions related to model training or model inference in the above embodiments.

This application also provides a computer program product that, when run on a computer, causes the computer to perform steps as performed by the aforementioned execution device, or causes the computer to perform steps as performed by the aforementioned training device.

This application also provides a computer-readable storage medium storing a program for signal processing, which, when run on a computer, causes the computer to perform steps as performed by the aforementioned execution device, or causes the computer to perform steps as performed by the aforementioned training device.

The execution device, training device, or terminal device provided in this application embodiment can specifically be a chip. The chip includes a processing unit and a communication unit. The processing unit can be, for example, a processor, and the communication unit can be, for example, an input/output interface, pins, or circuits. The processing unit can execute computer execution instructions stored in the storage unit to cause the chip within the execution device to execute the data processing method described in the above embodiments, or to cause the chip within the training device to execute the data processing method described in the above embodiments. Optionally, the storage unit is a storage unit within the chip, such as a register or cache. The storage unit can also be a storage unit located outside the chip within the wireless access device, such as read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, such as random access memory (RAM).

Specifically, please refer to Figure 11, which is a schematic diagram of a chip structure provided in an embodiment of this application. The chip can be represented as a neural network processor (NPU) 1100. The NPU 1100 is mounted as a coprocessor on the host CPU, and tasks are allocated by the host CPU. The core part of the NPU is the arithmetic circuit 1103, which is controlled by the controller 1104 to extract matrix data from the memory and perform multiplication operations.

In some implementations, the arithmetic circuit 1103 internally includes multiple processing engines (PEs). In some implementations, the arithmetic circuit 1103 is a two-dimensional pulsating array. The arithmetic circuit 1103 can also be a one-dimensional pulsating array or other electronic circuitry capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit 1103 is a general-purpose matrix processor.

For example, suppose we have an input matrix A, a weight matrix B, and an output matrix C. The arithmetic circuit retrieves the corresponding data of matrix B from the weight memory 1102 and caches it in each PE of the arithmetic circuit. The arithmetic circuit retrieves the data of matrix A from the input memory 1101 and performs matrix operations with matrix B. The partial result or the final result of the obtained matrix is stored in the accumulator 1108.

Unified memory 1106 is used to store input and output data. Weight data is directly transferred to weight memory 1102 via Direct Memory Access Controller (DMAC) 1105. Input data is also transferred to unified memory 1106 via DMAC.

BIU stands for Bus Interface Unit, which is used for the interaction between the AXI bus and the DMAC and the Instruction Fetch Buffer (IFB) 1109.

The Bus Interface Unit (BIU) 1110 is used by the instruction fetch memory 1109 to fetch instructions from external memory, and also by the memory access controller 1105 to fetch the original data of the input matrix A or the weight matrix B from external memory.

The DMAC is mainly used to move input data from external memory DDR to unified memory 1106, or to weight data to weight memory 1102, or to input data to input memory 1101.

The vector computation unit 1107 includes multiple processing units that further process the output of the computation circuit 1103 when needed, such as vector multiplication, vector addition, exponential operations, logarithmic operations, size comparisons, etc. It is mainly used for computation in non-convolutional/fully connected layers of neural networks, such as batch normalization, pixel-level summation, and upsampling of feature planes.

In some implementations, vector computation unit 1107 can store the processed output vector in unified memory 1106. For example, vector computation unit 1107 can apply a linear function, or a nonlinear function, to the output of computation circuit 1103, such as linear interpolation of feature planes extracted by convolutional layers, or, for example, a vector of accumulated values, to generate activation values. In some implementations, vector computation unit 1107 generates normalized values, pixel-level summed values, or both. In some implementations, the processed output vector can be used as activation input to computation circuit 1103, for example, for use in subsequent layers of the neural network.

The instruction fetch buffer 1109 connected to the controller 1104 is used to store the instructions used by the controller 1104;

Unified memory 1106, input memory 1101, weight memory 1102, and instruction fetch memory 1109 are all on-chip memories. External memory is proprietary to this NPU hardware architecture.

The processor mentioned above can be a general-purpose central processing unit, a microprocessor, an ASIC, or one or more integrated circuits used to control the execution of the above program.

It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims (26)

A data processing method, characterized in that the method includes: Obtain first data, which is obtained from the first input data through a machine learning model; The first data is subjected to non-uniform quantization processing to obtain the first compressed data, and the first compressed data is stored in the memory. The first compressed data is read from the memory, and the first compressed data is subjected to the inverse quantization process corresponding to the non-uniform quantization process to obtain the second data. The second data and the second input data are used to input into the machine learning model. The second input data is the data input into the machine learning model after the first input data. According to the method of claim 1, the first data is K data or V data. According to the method of claim 1 or 2, the first data comprises a plurality of data units, the first compressed data comprises a quantization value corresponding to each data unit, the data units of the first data comprise data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, the quantization value corresponding to the data interval in which each data unit is located is used as the quantization value corresponding to the data unit, wherein the numerical interval comprises a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval. The method according to any one of claims 1 to 3, characterized in that, the step of performing non-uniform quantization processing on the first data to obtain the first compressed data includes: Based on a preset mapping relationship, the data units of the first data are converted into corresponding quantized values to obtain the first compressed data; wherein, the mapping relationship includes multiple numerical intervals and the quantized value corresponding to each numerical interval, the data units of the first data include data in a first numerical range and data in a second numerical range, the data in the first numerical range is denser than the data in the second numerical range, the multiple numerical intervals include a first numerical interval and a second numerical interval, the first numerical interval belongs to the first numerical range, the second numerical interval belongs to the second numerical range, and the numerical width of the first numerical interval is smaller than the numerical width of the second numerical interval. The method according to any one of claims 1 to 4, characterized in that, the step of performing non-uniform quantization processing on the first data to obtain the first compressed data includes: The first data is subjected to a nonlinear transformation to obtain the transformed first data. The transformed first data is then subjected to uniform quantization. The method according to claim 5, characterized in that the method further comprises: Based on the quantization precision of the first data, determine the type of nonlinear function or the parameter values included in it when performing the nonlinear transformation. The method according to claim 5 or 6, characterized in that the first data is obtained from the first input data through the attention layer of a machine learning model, and the method further includes: Based on the position of the network layer containing the attention layer in the machine learning model, the type of nonlinear function or the parameter values included in the nonlinear transformation are determined. The method according to any one of claims 5 to 7, wherein the first data is obtained from the first input data through the target head in the attention layer of a machine learning model; the method further comprises: Based on the position of the target head in the attention layer, determine the type of nonlinear function or the parameter values included in the nonlinear transformation. The method according to any one of claims 5 to 8, characterized in that the method further comprises: Based on the generation interval between the first data and the latest data obtained by the machine learning model, determine the type of nonlinear function or the parameter values included when performing nonlinear transformation on the first data. The method according to any one of claims 5 to 9, characterized in that the method further comprises: Based on the data distribution of the first data, the category of the nonlinear function or the parameter values included in the nonlinear transformation of the first data are determined; the data distribution is indicated by the distribution statistics of the first data or by an identifier, wherein different identifiers correspond to data distributions with different characteristics. A data processing method, characterized in that the method includes: Obtain first data, which is calculated by a machine learning model based on first input data; The first data is transformed by a nonlinear function to obtain the transformed first data. The transformed first data is subjected to uniform quantization to obtain first compressed data, and the first compressed data is stored in a memory. The first compressed data is read from the memory, and the first compressed data is subjected to the inverse transformation process corresponding to the nonlinear transformation and the inverse quantization process corresponding to the uniform quantization process to obtain the second data. The second data and the second input data are used to input into the machine learning model. The second input data is the data input into the machine learning model after the first input data. The method according to claim 11, wherein the category of the nonlinear function or the numerical values of the included parameters are determined based on at least one of the following: The quantization precision of the first data; or, The position of the network that generated the first data within the machine learning model; or, The order in which the first data was generated in the machine learning model; or, The data distribution of the first data. A data processing apparatus, characterized in that the apparatus comprises: An acquisition module is used to acquire first data, which is obtained by a machine learning model based on first input data, and to read first compressed data from the memory; The processing module is used to perform non-uniform quantization processing on the first data to obtain the first compressed data, store the first compressed data in a memory, and perform inverse quantization processing on the first compressed data corresponding to the non-uniform quantization processing to obtain the second data. The second data and the second input data are used to input into the machine learning model. The second input data is the data input into the machine learning model after the first input data. The apparatus according to claim 13, wherein the first data comprises a plurality of data units, the first compressed data comprises a quantization value corresponding to each data unit, the data units of the first data comprise data within a first numerical range and data within a second numerical range, the data within the first numerical range being more densely packed than the data within the second numerical range, the quantization value corresponding to the data interval in which each data unit is located is used as the quantization value corresponding to the data unit, wherein the numerical interval comprises a first numerical interval and a second numerical interval, the first numerical interval belonging to the first numerical range, the second numerical interval belonging to the second numerical range, and the numerical width of the first numerical interval being smaller than the numerical width of the second numerical interval. The apparatus according to claim 13 or 14, wherein the processing module is specifically used for: The first data is subjected to a nonlinear transformation to obtain the transformed first data. The transformed first data is then subjected to uniform quantization. The apparatus according to claim 15, wherein the processing module is further configured to: Based on the quantization precision of the first data, determine the type of nonlinear function or the parameter values included in it when performing the nonlinear transformation. The apparatus according to claim 15 or 16, wherein the processing module is further configured to: Based on the position of the network layer containing the attention layer in the machine learning model, the type of nonlinear function or the parameter values included in the nonlinear transformation are determined. The apparatus according to any one of claims 15 to 17, wherein the first data is K data and V data obtained from the target head in the attention layer of a machine learning model based on the first input data; the processing module is further configured to: Based on the position of the target head in the attention layer, determine the type of nonlinear function or the parameter values included in the nonlinear transformation. The apparatus according to any one of claims 15 to 18, wherein the processing module is further configured to: Based on the generation interval between the first data and the latest data obtained by the machine learning model, the type of nonlinear function or the parameter values included in the nonlinear transformation are determined. The apparatus according to any one of claims 15 to 19, wherein the processing module is further configured to: Based on the data distribution of the first data, the category of the nonlinear function or the parameter values included in the nonlinear transformation are determined; the data distribution is indicated by the distribution statistics of the first data or by an identifier, wherein different identifiers correspond to data distributions with different characteristics. A data processing apparatus, characterized in that the apparatus comprises: An acquisition module is used to acquire first data, which is calculated based on first input data through a machine learning model, and to read the first compressed data from the memory; The processing module is configured to perform a nonlinear transformation on the first data using a nonlinear function to obtain the transformed first data; perform uniform quantization on the transformed first data to obtain the first compressed data, store the first compressed data in a memory, and perform inverse transformation processing corresponding to the nonlinear transformation and inverse quantization processing corresponding to the uniform quantization on the first compressed data to obtain second data. The second data and the second input data are used to input into the machine learning model, and the second input data is the data input into the machine learning model after the first input data. The apparatus according to claim 21, wherein the category of the nonlinear function or the numerical value of the included parameters is determined based on at least one of the following: The quantization precision of the first data; or, The position of the network that generated the first data within the machine learning model; or, The order in which the first data was generated in the machine learning model; or, The data distribution of the first data. A computer storage medium, characterized in that the computer storage medium stores one or more instructions, which, when executed by one or more computers, cause the one or more computers to perform the operation of the method according to any one of claims 1 to 12. A computer program product, characterized in that it includes computer-readable instructions that, when executed on a computer device, cause the computer device to perform the method as described in any one of claims 1 to 12. A system includes at least one processor and at least one memory; the at least one processor and the at least one memory are connected via a communication bus; The at least one memory is used to store code; The at least one processor is used to execute the code to perform the method as described in any one of claims 1 to 12. A chip includes a processor, characterized in that the processor is configured to implement the method as described in any one of claims 1 to 12.