Determination method for carbon emissions per unit of computing power in cloud computing task, and device

Abstract

Embodiments of the present invention provide a determination method for carbon emissions per unit of computing power in a cloud computing task, and a device. The method comprises: acquiring a neural network model corresponding to a cloud computing task; determining a computational graph corresponding to the neural network model, wherein the computational graph is composed of a plurality of computational graph elements, at least some of the plurality of computational graph elements are operable in parallel, and each computational graph element is any one of a node and an edge between nodes; on the basis of the parallel computational graph elements in the computational graph, determining a total computation amount corresponding to the neural network model during running; determining a total amount of carbon emissions corresponding to the neural network model; and on the basis of the total amount of carbon emissions and the total computation amount, determining carbon emissions per unit of computing power corresponding to the neural network model.

Description

[0001] Cross-Reference This disclosure claims priority to Chinese patent application No. 202410257294.X, filed with the Patent Office of China on March 6, 2024, entitled "Method and Apparatus for Determining Carbon Emissions per Unit of Computing Power of Cloud Computing Tasks," the entire contents of which are incorporated herein by reference. Technical Field This disclosure relates to the field of data processing technology, and more particularly to a method and apparatus for determining carbon emissions per unit of computing power of cloud computing tasks. Background: Against the backdrop of "carbon peak and carbon neutrality," data processing systems or data processing platforms face severe challenges in energy conservation and carbon emission reduction. Consequently, reducing carbon emissions per unit of computing power within data processing systems or platforms has become a current research hotspot. Currently, there is still no reliable and widely applicable method for calculating carbon emissions per unit of computing power, which hinders efforts to mitigate the serious harm caused by carbon emissions. SUMMARY OF THE INVENTION Embodiments of the present disclosure provide a method and apparatus for determining the carbon emissions per unit of computing power of a cloud computing task. These methods can accurately determine the carbon emissions per unit of computing power of a neural network model, and can then guide carbon emission operations based on the carbon emissions per unit of computing power, thereby facilitating carbon emission reduction. In a first aspect, embodiments of the present disclosure provide a method for determining the carbon emissions per unit of computing power of a cloud computing task, comprising: obtaining a neural network model corresponding to the cloud computing task; determining a computation graph corresponding to the neural network model, the computation graph comprising multiple computation graph elements, at least some of which can be operated in parallel, each computation graph element being any one of the following: a node or an edge between nodes; determining the total computational effort corresponding to the neural network model during execution based on the parallel computation graph elements included in the computation graph; determining the total carbon emissions corresponding to the neural network model; and determining the carbon emissions per unit of computing power corresponding to the neural network model based on the total carbon emissions and the total computational effort. In a second aspect, an embodiment of the present disclosure provides a device for determining carbon emissions per unit computing power of a cloud computing task, comprising: a first acquisition component for acquiring a neural network model corresponding to the cloud computing task; a first determination component for determining a computational graph corresponding to the neural network model, the computational graph being composed of multiple computational graph elements, at least a portion of the multiple computational graph elements being operable in parallel, and each computational graph element being any one of the following: a node, an edge between nodes; the first determination component is further used to determine the total computational amount corresponding to the neural network model during operation based on the parallel computational graph elements included in the computational graph; the first determination component is further used to determine the total carbon emissions corresponding to the neural network model; and a first processing component is used to determine the carbon emissions per unit computing power corresponding to the neural network model based on the total carbon emissions and the total computational amount.In a third aspect, embodiments of the present disclosure provide an electronic device comprising: a memory and a processor; wherein the memory is configured to store one or more computer instructions, wherein when executed by the processor, the one or more computer instructions implement the method for determining the carbon emissions per unit of computing power of a cloud computing task described in the first aspect. In a fourth aspect, embodiments of the present disclosure provide a computer storage medium configured to store a computer program, which, when executed by a computer, implements the method for determining the carbon emissions per unit of computing power of a cloud computing task described in the first aspect. In a fifth aspect, embodiments of the present disclosure provide a computer program product, comprising: a computer program, which, when executed by a processor of an electronic device, causes the processor to perform the steps of the method for determining the carbon emissions per unit of computing power of a cloud computing task described in the first aspect. In a sixth aspect, embodiments of the present disclosure provide a method for determining the total computational power of a cloud computing task, comprising: obtaining a neural network model for the cloud computing task; determining a computation graph corresponding to the neural network model, the computation graph consisting of multiple computation graph elements, at least some of which can be operated in parallel, each computation graph element being any one of the following: a node or an edge between nodes; and determining the total computational power corresponding to the neural network model during execution based on the parallel computation graph elements included in the computation graph. In a seventh aspect, embodiments of the present disclosure provide an apparatus for determining the total computational power of a cloud computing task, comprising: a second acquisition component for obtaining the neural network model for the cloud computing task; a second determination component for determining the computation graph corresponding to the neural network model, the computation graph consisting of multiple computation graph elements, at least some of which can be operated in parallel, each computation graph element being any one of the following: a node or an edge between nodes; and a second processing component for determining the total computational power corresponding to the neural network model during execution based on the parallel computation graph elements included in the computation graph. In an eighth aspect, embodiments of the present disclosure provide an electronic device, comprising: a memory and a processor; wherein the memory is configured to store one or more computer instructions, wherein when executed by the processor, the one or more computer instructions implement the method for determining the total computing power of a cloud computing task described in the sixth aspect. In a ninth aspect, embodiments of the present disclosure provide a computer storage medium, configured to store a computer program, wherein the computer program causes a computer to implement the method for determining the total computing power of a cloud computing task described in the sixth aspect when executed. In a tenth aspect, embodiments of the present disclosure provide a computer program product, comprising: a computer program, which, when executed by a processor of an electronic device, causes the processor to perform the steps of the method for determining the total computing power of a cloud computing task described in the sixth aspect.The method and device for determining the unit computing power carbon emissions of a cloud computing task provided by the embodiments of the present disclosure obtain a neural network model corresponding to the cloud computing task; determine a computation graph corresponding to the neural network model; and determine the total computational load corresponding to the neural network model during operation based on the parallel computation graph elements included in the computation graph. This effectively implements a computation graph-based statistical operation on the neural network computing power usage. By considering the impact of the parallel computation graph elements on the total computational load, the accuracy and reliability of the determination of the total computational load are effectively guaranteed. The total carbon emissions corresponding to the neural network model are then determined, and the unit computing power carbon emissions corresponding to the neural network model are determined based on the total carbon emissions and the total computational load, thereby effectively improving the accuracy and reliability of the determination of the unit computing power carbon emissions. The obtained unit computing power carbon emissions can be used as an indicator for uniformly measuring the computing power, energy consumption, and carbon emissions of intelligent computing scenarios. In this case, after obtaining the unit computing power carbon emissions, the unit computing power carbon emissions can be used to guide the carbon emissions work of the neural network model. This not only ensures the flexibility and reliability of the neural network model operation, but also helps reduce the carbon emissions of the system or platform, meeting the requirements of green environmental protection. This further improves the practicality of the method. BRIEF DESCRIPTION OF THE DRAWINGS To more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below represent some embodiments of the present disclosure. Those skilled in the art can derive other drawings based on these drawings without inventive effort.Figure 1 is a schematic diagram of the principle of a method for determining carbon emissions per unit computing power of a cloud computing task provided by an embodiment of the present disclosure; Figure 2 is a flowchart of a method for determining carbon emissions per unit computing power of a cloud computing task provided by an embodiment of the present disclosure; Figure 3 is a schematic diagram of a calculation graph provided by an embodiment of the present disclosure; Figure 4 is a schematic diagram of a calculation graph provided by an embodiment of the present disclosure; Figure 5 is a flowchart of determining the total computing amount corresponding to a neural network model during operation based on parallel calculation graph elements included in the calculation graph provided by an embodiment of the present disclosure; Figure 6 is a flowchart of determining the total carbon emissions corresponding to a neural network model provided by an embodiment of the present disclosure; Figure 7 is a schematic diagram of a calculation graph corresponding to an LSTM network model provided by an application embodiment of the present disclosure; Figure 8 is a flowchart of a method for determining the total computing amount of computing power provided by an embodiment of the present disclosure; Figure 9 is a schematic diagram of the structure of a device for determining carbon emissions per unit computing power provided by an embodiment of the present disclosure; Figure 10 is a schematic diagram of the structure of an electronic device corresponding to the device for determining carbon emissions per unit computing power provided by the embodiment shown in Figure 9; Figure 11 is a schematic diagram of the structure of a device for determining the total computing amount of computing power provided by an embodiment of the present disclosure; Figure 12 is a schematic diagram of the device for determining the total computing amount of computing power provided by an embodiment of the present disclosure; 1 is a schematic diagram of the structure of an electronic device corresponding to the device for determining the total amount of computing power provided in the embodiment shown in FIG. DETAILED DESCRIPTION OF THE EMBODIMENTS To further clarify the objectives, technical solutions, and advantages of the embodiments of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments represent only a portion of the embodiments of the present disclosure, and are not intended to be exhaustive. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present disclosure without inventive effort are within the scope of protection of the present disclosure. The terms used in the embodiments of the present disclosure are intended solely to describe specific embodiments and are not intended to limit the present disclosure. The singular forms "a," "an," and "the," as used in the embodiments of the present disclosure and the appended claims, are intended to include the plural forms, unless the context clearly indicates otherwise. "A plurality" generally includes at least two, but does not exclude the inclusion of at least one. It should be understood that the terms "and / or" as used herein are merely a description of an association relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B simultaneously, or B alone. Furthermore, in this document, Depending on the context, the phrases "if" and "if" as used herein may be interpreted as "upon..." or "when..." or "in response to determining..." or "in response to detecting." Similarly, depending on the context, the phrases "if it is determined" or "if (a stated condition or event) is detected" may be interpreted as "upon determination," "in response to determining," "upon detecting (a stated condition or event)," or "in response to detecting (a stated condition or event)." It should also be noted that the terms "comprise," "comprising," or any other variations thereof are intended to encompass a non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or system. Without further limitation, an element defined by the phrase "comprising a..." does not preclude the presence of additional identical elements in the product or system comprising the element. Furthermore, the sequence of steps in the following method embodiments is merely an example and not a strict limitation. To understand the specific implementation process of the technical solution in this embodiment, the following is a brief description of the relevant technologies: With the national carbon neutrality strategy and the "Eastern Data West Computing" project policy, "low-energy large models" and the coordinated development of computing power and electricity have become the focus of science and industry. Establishing a unit computing power carbon emission method suitable for intelligent computing scenarios has become a hot topic in research and engineering. Generally, the unit computing power carbon emission R can be defined as:

[0002] C

[0003] R =Q, where C represents total carbon emissions, measured in tons of carbon dioxide (CO2); Q represents the total computing power used during task execution, measured in floating-point operations. For data centers, computing power can be categorized into general-purpose computing power, intelligent computing power, network computing power, and storage computing power. However, due to the heterogeneity of computing tasks and the different computing power allocation logic, it is often difficult to calculate the various types of computing power required for a task. This makes reducing carbon emissions per unit of computing power during task processing challenging. In particular, intelligent computing models (typically neural network models), as one of the primary types of tasks handled by data centers, have garnered widespread attention regarding their computational carbon consumption. In related technology, researchers at the University of Massachusetts recently conducted a lifecycle assessment of a large-scale artificial intelligence model and found that the amount of carbon dioxide emitted by this process is roughly equivalent to the lifetime emissions of five mid-sized cars. Therefore, calculating the computing power used by neural network models and allocating computing resources to their operation is crucial for energy conservation and emission reduction in data centers. Specifically, for neural network models, the key to allocating computing resources lies in accurately analyzing the model's carbon footprint and fully leveraging its temporal and spatial flexibility. Analyzing the model's carbon footprint requires examining its topological structure. By analyzing the sparsity and data dependencies within the model layers, the model's response to various computing resources is determined, allowing the carbon emissions required at each stage of the model to be calculated. Currently, the carbon emissions corresponding to a neural network model can be determined by counting the model's total floating-point operations. However, this method fails to account for the potential parallelism within the model layers and, therefore, cannot accurately determine the model's performance on the graphics processor.

[0004] The computational power usage of highly parallelizable devices, such as Graphic Processing Units (GPUs) or Tensor Processing Units (TPUs), is measured. Furthermore, currently used open-source computational statistics methods only count computations within a layer, without considering the computational relationships between intermediate computational graph elements. This makes it difficult to flexibly provide unbiased statistical results in practice. Furthermore, industry-wide approaches to reducing the carbon emissions of neural network models primarily focus on reducing the size of the model itself, thereby lowering energy consumption during operation. Therefore, related technologies have proposed a method based on parameter quantization and pruning of deep neural networks to determine the carbon emissions corresponding to the neural network model. This approach reduces memory and computational consumption by reducing the number of bits in the model parameters, while sacrificing a small amount of accuracy. However, this approach focuses solely on the energy consumption of the model itself, ignoring the potential impact of the data center's computing resource allocation scheme on total energy consumption. Therefore, there is still room for improvement in achieving low-carbon goals. Furthermore, current computing resource allocation schemes do not adequately address the temporal flexibility of neural network operations (pausing or slowing down computing tasks during periods of high electricity and carbon prices) or spatial flexibility (deploying models across multiple devices in different locations to maximize computing resources while minimizing operating costs). Therefore, designing customized computing resource allocation schemes for different neural network models—enabling them to maximize computing resources within certain limits during periods of low electricity prices and carbon emissions, thereby completing more computing tasks—remains a significant challenge. To address the above technical issues, this embodiment provides a method, apparatus, and device for determining the carbon emissions per unit of computing power of a cloud computing task. Referring to FIG1 , the method for determining the carbon emissions per unit of computing power of a cloud computing task provided in this embodiment may be performed by an apparatus for determining the carbon emissions per unit of computing power of the cloud computing task. The apparatus for determining the carbon emissions per unit of computing power of the cloud computing task may be implemented as a terminal device, a personal computer, a tablet computer, a local server, or a cloud server. When the apparatus for determining the carbon emissions per unit of computing power of the cloud computing task is implemented as a cloud server, the method for determining the carbon emissions per unit of computing power of the cloud computing task may be performed in the cloud. The cloud may be equipped with multiple computing nodes (cloud servers), each of which has processing resources such as computing and storage. In the cloud, multiple computing nodes may be organized to provide a service, and a single computing node may also provide one or more services. The cloud may provide the service by providing an external service interface, which users invoke to access the corresponding service.Service interfaces include software development kits (SDKs) and application programming interfaces (APIs). The device for determining the carbon emissions per unit of computing power of a cloud computing task is communicatively connected to a client. The client is used by users to implement the operation of determining the carbon emissions per unit of computing power of a cloud computing task. The client can be any computing device with a certain data transmission capability. In specific implementations, the client can be a mobile phone, a personal computer (PC), a tablet computer, a configuration application, etc. Furthermore, the basic structure of the client may include at least one processor. The number of processors depends on the configuration and type of the client. The client may also include memory, which may be volatile, such as random access memory (RAM), non-volatile, such as read-only memory (ROM), flash memory, etc., or both types. The memory typically stores an operating system (OS), one or more application programs, and may also store program data. In addition to processing components and memory, the client also includes some basic configurations, such as a network card chip, an I / O bus, a display component, and some peripheral devices. Optionally, some peripheral devices may include, for example, a keyboard, a mouse, an input pen, a printer, and the like. Other peripheral devices are well known in the art and are not described in detail here. A device for determining the carbon emissions per unit of computing power of a cloud computing task refers to a device that can provide operations for determining the carbon emissions per unit of computing power of a cloud computing task in a network virtual environment. It typically refers to a device that utilizes a network to perform information planning and determine the carbon emissions per unit of computing power of a cloud computing task. In physical implementation, the device for determining the carbon emissions per unit of computing power of a cloud computing task can be any device capable of providing computing services, responding to requests for determining the carbon emissions per unit of computing power of a cloud computing task, and performing the determination operation based on the request. Examples include cluster servers, conventional servers, cloud servers, cloud hosts, virtual centers, and the like. The device for determining the carbon emissions per unit of computing power of a cloud computing task primarily includes a processor, hard disk, memory, and system bus, similar to a general computer architecture.In the above-described embodiment, a client establishes a network connection with the apparatus for determining carbon emissions per unit computing power of a cloud computing task. This network connection can be a wireless or wired network connection. In response to the client establishing a communication connection with the apparatus for determining carbon emissions per unit computing power of a cloud computing task, the mobile network standard can be any one of 2G (GSM), 2.5G (GPRS), 3G (WCDMA, TD-SCDMA, CDMA2000, UTMS), 4G (LTE), 4G+ (LTE+), Wi-Fi Max, 5G, 6G, etc. In this embodiment of the present application, the client is configured to generate or obtain pending cloud computing tasks. Specifically, the client can display a human-computer interaction interface, obtain execution operations entered by a user in the human-computer interaction interface, and generate or obtain pending cloud computing tasks based on the execution operations. To determine the carbon emissions per unit computing power of the cloud computing task, the pending cloud computing task can be sent to the apparatus for determining carbon emissions per unit computing power of the cloud computing task. A device for determining the carbon emissions per unit computing power of a cloud computing task is used to obtain a cloud computing task sent by a client. To implement the operation of determining the carbon emissions per unit computing power in intelligent computing scenarios, after obtaining the cloud computing task, the corresponding neural network model can be obtained through the cloud computing task. Since the carbon emissions per unit computing power are related to the architecture or structure of the neural network model, to accurately determine the carbon emissions per unit computing power of the neural network model, after obtaining the neural network model corresponding to the cloud computing task, the neural network model can be analyzed and processed to determine the computation graph corresponding to the neural network model. Since the computation graph is composed of multiple computation graph elements, at least some of the multiple computation graph elements can be operated in parallel. Each computation graph element is any one of the following: a node or an edge between nodes. The nodes are used to identify computation relationships, and the edges between nodes are used to identify node states or tensor states. Specifically, a node can include at least one of the following: a transfer function, an accumulation operation, a multiplication operation, a transposition operation, etc., and an edge can include at least one of the following: a model parameter, an input matrix, etc. Because at least some of the multiple computational graph elements in a computational graph can operate in parallel, and these parallel computational graph elements can directly affect the accuracy of carbon emissions per unit of computing power, after obtaining the computational graph corresponding to the neural network model, the total computational effort corresponding to the neural network model during operation can be determined based on the parallel computational graph elements included in the computational graph. This effectively ensures the accuracy and reliability of the determination of the total computational effort.Furthermore, to accurately determine the carbon emissions per unit of computing power, it is necessary not only to accurately determine the total computational effort corresponding to the neural network model but also the total carbon emissions corresponding to the neural network model. The carbon emissions per unit of computing power corresponding to the neural network model can then be determined based on the determined total carbon emissions and total computational effort. In some examples, carbon emissions per unit of computing power = total carbon emissions / total computational effort. This effectively implements computational graph-based statistical operations for neural network computing power usage. Furthermore, by considering the impact of parallel computational graph elements on the total computational effort, the accuracy and reliability of the total computational effort determination are effectively guaranteed, improving the accuracy and reliability of the carbon emissions per unit of computing power determination. The obtained carbon emissions per unit of computing power can be used as a unified metric for measuring computing power, energy consumption, and carbon emissions in intelligent computing scenarios. Specifically, the carbon emissions per unit of computing power can be used to guide the carbon emissions assessment of the neural network model. This not only ensures the flexibility of the neural network model operation but also helps reduce the carbon emissions of the system or platform, further enhancing the practicality of this method. The following describes some embodiments of the present disclosure in detail with reference to the accompanying drawings. Unless there is a conflict between the various embodiments, the following embodiments and their features may be combined with each other. Furthermore, the sequence of steps in the following method embodiments is provided for illustrative purposes only and is not a strict limitation. Figure 2 is a flowchart illustrating a method for determining carbon emissions per unit computing power of a cloud computing task, provided in an embodiment of the present disclosure. Referring to Figure 2 , this embodiment provides a method for determining carbon emissions per unit computing power of a cloud computing task. The method may be performed by an apparatus for determining carbon emissions per unit computing power of a cloud computing task. It is understood that the apparatus for determining carbon emissions per unit computing power of a cloud computing task may be implemented as software or a combination of software and hardware. Specifically, when the apparatus for determining carbon emissions per unit computing power of a cloud computing task is implemented as hardware, it may be various electronic devices capable of determining carbon emissions per unit computing power of a cloud computing task, including but not limited to tablet computers, personal computers (PCs), servers, and the like. When the apparatus for determining carbon emissions per unit computing power of a cloud computing task is implemented as software, it may be installed in the electronic devices listed above. Based on the aforementioned apparatus for determining carbon emissions per unit computing power of a cloud computing task, the method for determining carbon emissions per unit computing power of a cloud computing task in this embodiment may include the following steps: Step S201: Obtaining a neural network model corresponding to the cloud computing task. Step S202: Determining a computation graph corresponding to the neural network model. The computation graph is composed of multiple computation graph elements, at least some of which can operate in parallel. Each computation graph element is any one of the following: a node, or an edge between nodes.Step S203: Based on the parallel computation graph elements included in the computation graph, determine the total computational effort corresponding to the neural network model during operation. Step S204: Determine the total carbon emissions corresponding to the neural network model. Step S205: Based on the total carbon emissions and the total computational effort, determine the unit computing power carbon emissions corresponding to the neural network model. The specific implementation principles and results of each of the above steps are detailed below: Step S201: Obtain the neural network model corresponding to the cloud computing task. When a user has a requirement to determine the carbon emissions per unit of computing power for a cloud computing task, a device for determining the carbon emissions per unit of computing power (hereinafter referred to as a "determination device") can obtain a neural network model corresponding to the cloud computing task. In some instances, the neural network model corresponding to the cloud computing task can be obtained through human-computer interaction. At this time, obtaining the neural network model corresponding to the cloud computing task can include: displaying multiple pending cloud computing tasks through a human-computer interaction interface; obtaining a selection operation input by the user on any one of the multiple pending cloud computing tasks, and determining the cloud computing task corresponding to the selection operation as the neural network model corresponding to the cloud computing task, thereby effectively achieving the ability to stably obtain the neural network model corresponding to the cloud computing task based on human-computer interaction. In other examples, the neural network model corresponding to the cloud computing task can be obtained not only through human-computer interaction, but also through a client or a third device. In this case, obtaining the neural network model corresponding to the cloud computing task may include: obtaining a client or a third device in communication with the determination device, wherein the client or the third device stores the neural network model corresponding to the cloud computing task. The client or the third device actively or passively obtains the neural network model corresponding to the cloud computing task. This effectively ensures the flexibility and reliability of determining the neural network model corresponding to the cloud computing task. Step S202: Determine a computation graph corresponding to the neural network model. The computation graph is composed of multiple computation graph elements, at least some of which can be operated in parallel. Each computation graph element is any one of the following: a node, an edge between nodes.Since different neural network models have different model architectures, neural network models with different model architectures will generate different computational loads. Therefore, in order to accurately calculate the total computational load corresponding to the neural network model during operation, after obtaining the neural network model, the neural network model can be analyzed and processed to determine the computational graph corresponding to the neural network model. In some instances, the computational graph can be determined by the model type and model architecture of the neural network model. At this time, determining the computational graph corresponding to the neural network model can include: obtaining the model type and model architecture corresponding to the neural network model; and determining the computational graph corresponding to the neural network model based on the model type and model architecture. As shown in FIG3 , the determined computation graph may be composed of multiple computation graph elements. Each computation graph element includes any one of the following: a node and an edge between nodes. Nodes are used to identify computation relationships, and edges between nodes are used to identify node states or tensor states. Specifically, a node may include at least one of the following: a transfer function, an accumulation operation, a multiplication operation, a transposition operation, etc., and an edge may include at least one of the following: a model parameter, an input matrix, etc. Those skilled in the art may flexibly configure or adjust the specific type of computation graph elements based on the specific application scenario or application requirements of the neural network model. Furthermore, the multiple computation graph elements in the computation graph may include serial computation graph elements that operate serially and parallel computation graph elements that can operate in parallel. That is, the computation graph is composed of multiple computation graph elements, and at least a portion of the multiple computation graph elements can operate in parallel. Because the number of computational graph elements that can be operated in parallel can directly affect the total computational workload of the neural network model during operation, after determining the computational graph corresponding to the neural network model, the parallel computational graph elements in the computational graph can be identified. The parallel computational graph elements in the computational graph can be obtained by analyzing and processing the computational graph using a pre-trained machine learning model. Alternatively, the parallel computational graph elements can be determined based on the computational layer and data dependency relationships of the computational graph elements. In this case, the method in this embodiment may further include: determining, based on the computational graph, the computational layer to which each computational graph element belongs and whether a data dependency relationship exists between any two computational graph elements; if two computational graph elements belong to the same computational layer and no data dependency relationship exists between the two computational graph elements, determining that the two computational graph elements are parallel computational graph elements; if the two computational graph elements belong to different computational layers or a data dependency relationship exists between the two computational graph elements, determining that the two computational graph elements are non-parallel computational graph elements.For example, taking the computation graph of an artificial neural network (ANN) in the Pytorch framework as an example, as shown in FIG4 , if the accumulation operation device, transposition operation device, and weight parameter device in the computation graph belong to the same computation layer and there is no data dependency between the accumulation operation device, transposition operation device, and weight parameter device, then the accumulation computation graph element, transposition computation graph element, and weight computation graph element can be determined to be parallel computation graph elements. If the bias computation graph element and activation function computation graph element in the computation graph belong to different computation layers, or if there is a data dependency between the accumulation computation graph element and the transposition computation graph element, then the bias computation graph element, activation function computation graph element, accumulation computation graph element, and transposition computation graph element can all be determined to be non-parallel computation graph elements, thereby effectively achieving accurate identification of parallel computation graph elements in the computation graph. Step S203: Based on the parallel computation graph elements included in the computation graph, determine the total computation amount corresponding to the neural network model during operation. Since the total computational load corresponding to the neural network model during operation is directly related to the parallel computation graph elements in the computation graph, after obtaining the computation graph, the parallel computation graph elements included in the computation graph can be analyzed and processed to determine the total computational load corresponding to the neural network model during operation. In some instances, the total computational load can be determined using a pre-defined mapping relationship. In this case, determining the total computational load corresponding to the neural network model during operation based on the parallel computation graph elements included in the computation graph can include: obtaining a preset mapping table for determining computational load, the preset mapping table including a mapping relationship between the computation graph type, the number of parallel computation graph elements, and computational load; determining the computation graph type and the number of parallel computation graph elements; and searching the preset mapping table based on the computation graph type and the number of parallel computation graph elements, thereby stably obtaining the total computational load corresponding to the neural network model during operation. Step S204: Determining the total carbon emissions corresponding to the neural network model.In order to accurately determine the carbon emissions per unit computing power corresponding to the neural network model, it is necessary not only to obtain the total computing amount corresponding to the neural network model during operation, but also to determine the total carbon emissions corresponding to the neural network model. In some instances, the total carbon emissions corresponding to the neural network model can be determined based on the model type and running time of the neural network model. In this case, determining the total carbon emissions corresponding to the neural network model may include: obtaining the model type and running time of the neural network model; and determining the total carbon emissions corresponding to the neural network model based on the model type and running time. Specifically, the total carbon emissions corresponding to the neural network model can be determined using a mapping relationship, model type, and running time. This effectively ensures the accuracy and reliability of the determination of the total carbon emissions. It should be noted that the execution order between step S204 and steps S202-S203 in this embodiment is not limited to the implementation methods defined in the above embodiments. For example, step S204 may be executed before steps S202-S203, or step S204 may be executed simultaneously with steps S202-S203. Those skilled in the art may flexibly adjust the execution order of steps S204 and steps S202-S203 based on specific application scenarios or application requirements. Step S205: Based on the total carbon emissions and the total computational load, the unit computing power carbon emissions corresponding to the neural network model are determined. After obtaining the total carbon emissions and the total computational load, the total carbon emissions and the total computational load can be analyzed and processed. In some examples, the unit computing power carbon emissions corresponding to the neural network model can be determined as the ratio of the total carbon emissions to the total computational load, i.e., unit computing power carbon emissions R = total carbon emissions C / total computational load Q. This allows accurate determination of the unit computing power carbon emissions corresponding to the neural network model. To improve the practicality of this method, after determining the carbon emissions per unit of computing power corresponding to the neural network model, the operation of the neural network model can be guided based on the carbon emissions per unit of computing power. For example, when the carbon emissions per unit of computing power is greater than or equal to a preset threshold, the neural network model can be controlled to suspend operation; when the carbon emissions per unit of computing power is less than the preset threshold, the neural network model can be kept running, and so on.Furthermore, since the unit computing power carbon emissions determined by the neural network model in different time periods may vary, the operation of the neural network model can be guided not only directly based on the unit computing power carbon emissions, but also by obtaining a unit computing power carbon emissions curve within a preset time period and then guiding the operation of the neural network model based on the unit computing power carbon emissions curve. In this case, the method in this embodiment may include: obtaining multiple unit computing power carbon emissions within the preset time period; obtaining a unit computing power carbon emissions curve based on the multiple unit computing power carbon emissions and the time information corresponding to each unit computing power carbon emissions; then predicting the unit computing power carbon emissions of the neural network model at a future time based on the unit computing power carbon emissions curve, and adjusting and controlling the operation of the neural network model based on the predicted unit computing power carbon emissions to ensure the quality and efficiency of data processing based on the neural network model. In addition, this embodiment can also control the operation of the neural network model at the current moment based on the carbon emission curve per unit computing power predicted at a future moment. Specifically, the method in this embodiment can also include: obtaining the carbon emission curve per unit computing power predicted by the neural network model within a preset time period in the future (which can be the next day, the next week, etc.); then, based on the carbon emission curve per unit computing power and the computational load of the neural network model, obtaining the carbon emission per unit computing power corresponding to the neural network model at different moments. Then, the operation of the neural network model can be adjusted and controlled based on the carbon emission per unit computing power corresponding to the neural network model at different moments. This not only reduces the carbon emission per unit computing power, but also ensures the quality and efficiency of data processing operations.This embodiment provides a method for determining the carbon emissions per unit computing power of a cloud computing task. The method obtains a neural network model corresponding to the cloud computing task, determines a computation graph corresponding to the neural network model, and determines the total computational load corresponding to the neural network model during operation based on the parallel computation graph elements included in the computation graph. This effectively implements a computation graph-based statistical operation on the neural network computing power usage. By taking into account the impact of the parallel computation graph elements on the total computational load, the accuracy and reliability of the determination of the total computational load are effectively guaranteed. The total carbon emissions corresponding to the neural network model are then determined, and the carbon emissions per unit computing power corresponding to the neural network model are determined based on the total carbon emissions and the total computational load. This effectively improves the accuracy and reliability of the determination of the carbon emissions per unit computing power. The obtained carbon emissions per unit computing power can be used as an indicator for uniformly measuring the computing power, energy consumption, and carbon emissions of intelligent computing scenarios. Once the carbon emissions per unit computing power are obtained, the carbon emissions per unit computing power can be used to guide the carbon emissions assessment of the neural network model. This not only ensures the flexibility and reliability of the neural network model operation, but also helps reduce the carbon emissions of the system or platform, meeting green environmental protection requirements and further improving the practicality of the method. FIG5 is a flow chart illustrating a method for determining the total computational load corresponding to a neural network model during operation based on parallel computation graph elements included in a computation graph, according to an embodiment of the present disclosure. Based on the above embodiment, and with reference to FIG5 , the total computational load corresponding to a neural network model during operation can be determined not only by a preset mapping relationship but also by correcting the computational load corresponding to the neural network model during operation based on the parallel computation graph elements in the computation graph. In this case, determining the total computational load corresponding to a neural network model during operation based on the parallel computation graph elements in the computation graph may include the following: Step S501: Obtaining the original computational load corresponding to the neural network model during operation based on the computation graph. Step S502: Correcting the original computational load using the parallel computation graph elements in the computation graph to obtain the total computational load corresponding to the neural network model during operation. Specifically, after obtaining the computation graph, the computation graph can be analyzed and processed to obtain the original computational load corresponding to the operation of the neural network model. In some instances, the original computational load can be obtained based on a pre-trained machine learning model, and the computation graph can be analyzed and processed using the machine learning model to obtain the original computational load corresponding to the operation of the neural network model.In other instances, the raw computational load corresponding to the operation of the neural network model can be determined not only through a machine learning model but also based on the forward propagation and backpropagation processes of the neural network model. In this case, obtaining the raw computational load corresponding to the operation of the neural network model based on the computation graph may include: traversing the computation graph to obtain a first computational load corresponding to one forward propagation operation and a second computational load corresponding to one backpropagation operation of the neural network model; determining the number of training stages and inference stages corresponding to the neural network model; and determining the raw computational load corresponding to the operation of the neural network model based on the number of training stages, the number of inference stages, the first computational load, and the second computational load. For a neural network model, different data processing operations correspond to different computational loads. For example, the computational load of a neural network model after a forward propagation operation is different from the computational load after a backpropagation operation, and the computational loads after the forward propagation operation and the backpropagation operation are correlated with the total computational load of the neural network model. Therefore, in order to accurately determine the computational load corresponding to the neural network model during operation, after obtaining the computation graph, the computation graph can be traversed to obtain a first computational load corresponding to one forward propagation operation of the neural network model and a second computational load corresponding to one backward propagation operation. In some instances, the first computational load and the second computational load can be determined based on the model type or model structure of the neural network model, or the first computational load and the second computational load can be determined by a preset operation corresponding to the forward propagation operation and a preset operation corresponding to the backward propagation operation, respectively. Traversing the computation graph to obtain the first computational load corresponding to one forward propagation operation of the neural network model can include: traversing the computation graph to obtain input computation graph elements, computation graph elements corresponding to model parameters, matrix parameters corresponding to matrix multiplication operations, the number of matrix multiplication operations, and the number of operations of the computation graph elements corresponding to the model parameters, included in one forward propagation operation of the neural network model; and based on the input computation graph elements, the computation graph elements corresponding to the model parameters, the matrix parameters corresponding to the matrix multiplication operations, the number of matrix multiplication operations, and the number of operations of the computation graph elements corresponding to the model parameters, Determine a first computational amount corresponding to one forward pass operation of the neural network model.For example, after obtaining the computation graph, the computation graph can be traversed and processed to obtain the input computation graph element π, the computation graph element π corresponding to the model parameter, the number of matrix multiplication operations π, and the number of operations π for the computation graph element corresponding to the model parameter. When the neural network model includes a matrix multiplication operation between a matrix gate and a matrix π*k, the matrix parameters π, π, and κ corresponding to the matrix multiplication operation can be obtained. After obtaining the computation graph element π corresponding to the model parameter, the matrix parameters m, π, and k corresponding to the matrix multiplication operation, the number of matrix multiplication operations π, and the number of operations π for the computation graph element corresponding to the model parameter, the formula A=π=0 (π / π / (2π+3)*π+π / π) can be used to obtain the first computation amount corresponding to one forward propagation operation of the neural network model. This effectively ensures that the first computation amount A corresponding to one forward propagation operation of the neural network model is accurately determined.Similarly, traversing the computational graph to obtain the second computational cost corresponding to one backpropagation operation of the neural network model may include: traversing the computational graph to obtain the input computational graph elements, the computational graph elements corresponding to the model parameters, the transfer relationships, the matrix parameters corresponding to the matrix multiplication operations, the number of matrix multiplication operations, the first operation count of the computational graph elements corresponding to the model parameters, and the second operation count of the transfer relationships; based on the input computational graph elements, the computational graph elements corresponding to the model parameters, the transfer relationships, the matrix parameters corresponding to the matrix multiplication operations, the number of matrix multiplication operations, the first operation count of the computational graph elements corresponding to the model parameters, and the second operation count of the transfer relationships, determine the second computational cost corresponding to one backpropagation operation of the neural network model. For example, after obtaining the computational graph, the input computational graph elements %, the computational graph elements corresponding to the model parameters %, the transfer relationship 9, the number of matrix multiplication operations 0, the operation count of the computational graph elements corresponding to the model parameters 勿, and the second operation count of the transfer relationships R can be obtained through traversing and processing the computational graph; when the neural network model includes a matrix multiplication operation between matrix m * 71 and matrix 71 * k, the matrix parameters 771, 71, k corresponding to the matrix multiplication operation can be obtained. After obtaining the input computational graph elements %, the computational graph elements corresponding to the model parameters %, the transfer relationship 9, the number of matrix multiplication operations 0, the operation count of the computational graph elements corresponding to the model parameters 0E, the second operation count of the transfer relationships R, and the matrix parameters 771, 71, k corresponding to the matrix multiplication operation, the formula B = &=0(2店冲1(2化 + 3) *。 + b® + q (10,000) to obtain the second computational amount corresponding to a single back-propagation operation of the neural network model, thereby effectively ensuring the accurate determination of the second computational amount B corresponding to a single back-propagation operation of the neural network model. It should be noted that the first computational amount A and the second computational amount B can be determined not only by the above formula, but also by analyzing and processing various parameters using other algorithms or formulas. For example, the first computational amount can be obtained by analyzing and processing the input computational graph elements, the computational graph elements corresponding to the model parameters, the matrix parameters corresponding to the matrix multiplication operation, the number of matrix multiplication operations, and the number of operations of the computational graph elements corresponding to the model parameters using a first preset algorithm or a first preset formula. Alternatively, the second computational amount can be obtained by analyzing and processing the input computational graph elements, the computational graph elements corresponding to the model parameters, the transfer relationship, the matrix parameters corresponding to the matrix multiplication operation, the number of matrix multiplication operations, the first number of operations of the computational graph elements corresponding to the model parameters, and the second number of operations corresponding to the transfer relationship using a second preset algorithm or a second preset formula. As long as the accuracy and reliability of obtaining the first computational amount A and the second computational amount B can be guaranteed, no further details will be given here. After obtaining the first and second computational loads, since the neural network model may correspond to different numbers of training stages and inference stages in different application scenarios, in order to accurately determine the computational load corresponding to the neural network model during operation, the number of training stages and inference stages corresponding to the neural network model can be determined. In some instances, the number of training stages and inference stages can be flexibly configured or adjusted based on the application scenario or application requirements. After determining the number of training stages, the number of inference stages, the first and second computational loads corresponding to the neural network model, the number of training stages, the number of inference stages, the first and second computational loads can be analyzed and processed to determine the original computational load corresponding to the neural network model during operation. In some instances, determining the original amount of computation corresponding to the neural network model during operation based on the number of training stages, the number of inference stages, the first amount of computation, and the second amount of computation may include: obtaining a computation sum value between the first amount of computation and the second amount of computation; determining a first product value between the computation sum and the number of training stages, and a second product value between the first amount of computation and the number of inference stages; and determining the sum of the first product value and the second product value as the original amount of computation corresponding to the neural network model during operation.For example, when the number of training stages is train, the number of inference stages is infer, and the first computation amount is A and the second computation amount is B, the computation amount sum A+B of the first and second computation amounts can be obtained. Then, a first product value, train*(A+B), between the computation amount sum and the number of training stages and a second product value, infer*k, between the first computation amount and the number of inference stages can be determined. The sum of the first and second product values ​​can then be determined as the raw computation amount corresponding to the neural network model during operation. Specifically, the raw computation amount corresponding to the neural network model during operation can be obtained using the following formula: Q=infer*A+train*(A+B). This effectively ensures the accuracy and reliability of determining the raw computation amount. It should be noted that the raw computation amount corresponding to the neural network model during operation can be determined not only using the above formula but also by analyzing and processing the number of training stages, the number of inference stages, the first computation amount, and the second computation amount using a preset algorithm or a preset comparison table. This also ensures flexibility and reliability in determining the second computation amount. Since the original computational load obtained above does not take into account the impact of the parallel computation graph elements included in the computation graph, to ensure accurate and reliable determination of the total computational load, the parallel computation graph elements included in the computation graph can be used to modify the original computation graph, thereby obtaining the total computational load corresponding to the neural network model during operation. In some instances, the modification operation can be implemented not only through a preset algorithm or a preset lookup table, but also by obtaining a modification coefficient. In this case, modifying the original computational load using the parallel computation graph elements included in the computation graph to obtain the total computational load corresponding to the neural network model during operation may include: determining a modification coefficient for modifying the original computational load based on the parallel computation graph elements included in the computation graph and the original computational load; and obtaining the total computational load corresponding to the neural network model during operation based on the modification coefficient and the original computational load.Specifically, since the parallel computation graph elements included in the computation graph can affect the computational load corresponding to the neural network model, in order to accurately determine the computational load corresponding to the neural network model during operation, a correction coefficient for correcting the original computational load can be determined based on the parallel computation graph elements included in the computational load. In some instances, the correction coefficient can be obtained through a preset algorithm or a preset comparison table. At this time, determining the correction coefficient for correcting the original computational load based on the parallel computation graph elements included in the computation graph and the original computational load may include: obtaining a preset algorithm or a preset comparison table, inputting the parallel computation graph elements included in the computation graph and the original computational load into the machine learning model, and obtaining the correction coefficient for correcting the original computational load. This effectively ensures the accuracy and reliability of determining the correction coefficient. In other instances, the correction coefficient can be obtained not only through a preset algorithm or a preset comparison table, but also based on the parallel computing amount corresponding to the parallel computing graph element. In this case, based on the parallel computing graph elements and the original computing amount included in the computing graph, determining the correction coefficient for correcting the original computing amount may include: obtaining the parallel computing amount corresponding to the parallel computing graph elements in the computing graph and the weight coefficient corresponding to the parallel computing amount; determining the non-parallel computing amount corresponding to the neural network model based on the parallel computing amount and the original computing amount; and determining the correction coefficient for correcting the original computing amount based on the non-parallel computing amount, the parallel computing amount, the weight coefficient, and the original computing amount. Since the computational load corresponding to the neural network model is related to the computational load corresponding to the parallel computation graph elements in the computation graph, in order to accurately determine the correction coefficient used to correct the original computational load, after obtaining the parallel computation graph elements in the computation graph, the parallel computation load (S - SQ) corresponding to the parallel computation graph elements in the computation graph and the weight coefficient KK corresponding to the parallel computation load can be obtained. Specifically, the parallel computation load can be determined by the element type or number of the parallel computation graph elements in the computation graph, and the weight coefficient corresponding to the parallel computation load can be obtained through human-computer interaction or pre-configuration. Since the computation graph may include not only parallel computation graph elements but also non-parallel computation graph elements, and since the computational load of the neural network model is related not only to the parallel computation graph elements but also to the non-parallel computation graph elements, in order to accurately determine the correction coefficient used to correct the original computational load, after obtaining the parallel computation load, the parallel computation load and the original computation load can be analyzed and processed to determine the non-parallel computation load corresponding to the neural network model. Specifically, the parallel computation load corresponding to the parallel computation graph elements in the computation graph is (S - S). K), when the original computation amount is S, the nonlinearity corresponding to the neural network model can be determined by the parallel computation amount and the original computation amount. After obtaining the non-parallel computation amount &, the parallel computation amount (s - SQ), the weight coefficient 6K, and the original computation amount s, the non-parallel computation amount, the parallel computation amount, the weight coefficient, and the original computation amount can be analyzed and processed. In some instances, the correction coefficient can be obtained by analyzing and processing the non-parallel computation amount, the parallel computation amount, the weight coefficient, and the original computation amount using a pre-trained machine learning model, thereby determining the correction coefficient for correcting the original computation amount. In other instances, the correction coefficient can be obtained not only by analyzing and processing the non-parallel computation amount, the parallel computation amount, the weight coefficient, and the original computation amount using a pre-trained machine learning model, but also by obtaining the correction coefficient for correcting the original computation amount using a preset algorithm formula. In this case, determining the correction coefficient for correcting the original computation amount based on the non-parallel computation amount, the parallel computation amount, the weight coefficient, and the original computation amount can include: obtaining a learnable parameter for determining the correction coefficient; determining a product value of the parallel computation amount and the weight coefficient, a sum of the product value and the computation amount of the non-parallel computation amount, and a sum of the product value and the computation amount of the non-parallel computation amount; The ratio between the sum of the computational load and the original computational load is used; the exponent of the ratio is used as the base and the learnable parameter is used as the exponent to determine a correction coefficient for correcting the original computational load. Specifically, to accurately determine the correction coefficient for correcting the original computational load, a learnable parameter for determining the correction coefficient can be obtained or designed. The learnable parameter for determining the correction coefficient can be stored in a preset area or a preset device, and the learnable parameter for determining the correction coefficient can be obtained by accessing the preset area or the preset device. The learnable parameter can be yK. The learnable parameters y and K can be obtained through separate regression tests. After the neural network model training operation is completed, the learnable parameters y and K can be fixed. After obtaining the learnable parameter for determining the correction coefficient, the product value of the parallel computational load and the weight coefficient can be determined as y<(S - SQ, the product value and the sum of the non-parallel computational load) + yK(S - SQ, the ratio between the sum of the computational load and the original computational load) + y<(S - SK). The ratio is then used as the base and the original value. The exponent of the exponent can be learned and used as the correction coefficient for correcting the original calculation amount, which effectively ensures the accuracy of the correction coefficient. After obtaining the correction coefficient and the original computational load, the correction coefficient and the original computational load can be analyzed and processed to obtain the total computational load corresponding to the operation of the neural network model. In some examples, the product of the correction coefficient and the original computational load is determined as the total computational load corresponding to the operation of the neural network model, that is, the total computational load C = ctK (S) * S, where S is the original computational load and the correction coefficient is a K(S). In this embodiment, the raw computational load corresponding to the neural network model during operation is obtained based on the computation graph. A correction coefficient for correcting the raw computational load is determined based on the parallel computation graph elements included in the computation graph and the raw computational load. The total computational load corresponding to the neural network model during operation can then be obtained based on the correction coefficient and the raw computational load, thereby effectively ensuring the accuracy and reliability of the determination of the total computational load. Figure 6 is a schematic flow diagram of determining the total carbon emissions corresponding to a neural network model according to an embodiment of the present disclosure. Based on any of the above embodiments, referring to Figure 6, the total carbon emissions corresponding to the neural network model can be determined not only based on the model type and runtime of the neural network model, but also based on a pre-trained reinforcement learning model. In this case, determining the total carbon emissions corresponding to the neural network model may include the following: Step S601: Obtaining a pre-trained reinforcement learning model for determining a resource allocation strategy. The reinforcement learning model is trained using the resource allocation strategy of the neural network model as a decision variable and reducing carbon emissions within a preset time period as an optimization objective. Because neural network models can achieve different carbon reduction effects under different resource allocation scenarios, a reinforcement learning model is pre-configured to determine resource allocation strategies in order to fully leverage the temporal and spatial flexibility of neural network models to achieve low-carbon goals. A neural network's resource allocation strategy involves allocating the neural network to different numbers of devices for execution at different times. The reinforcement learning model can be used to determine resource allocation strategies for different application scenarios. For example, when the combined carbon emissions and electricity costs within a preset time period are low, the neural network model's resource allocation strategy can be determined to be a tensor parallel execution strategy. When carbon emissions within a preset time period are high, the neural network model's resource allocation strategy can be determined to be a non-tensor parallel execution strategy, and so on. This reinforcement learning model can be used to adjust the neural network model's computing resource allocation strategy in different application scenarios, ensuring that the neural network model maximizes the use of idle resources for data processing during periods of low electricity and carbon prices, while minimizing computing resource consumption during periods of low prices. This effectively reduces the neural network model's carbon emissions.Furthermore, the reinforcement learning model can be obtained not only by training with the resource allocation strategy of the neural network model as a decision variable and reducing carbon emissions within a preset time period as an optimization objective, but also by training with the resource allocation strategy of the neural network model as a decision variable and reducing the combined loss of carbon emissions, electricity costs, and service quality within a preset time period as an optimization objective. Alternatively, the reinforcement learning model can be obtained by training with the resource allocation strategy of the neural network model as a decision variable and reducing the combined loss of carbon emissions and electricity costs within a preset time period as an optimization objective. Alternatively, the reinforcement learning model can be obtained by training with the resource allocation strategy of the neural network model as a decision variable and reducing the combined loss of carbon emissions and service quality within a preset time period as an optimization objective. This not only allows for flexible adjustment of the operation strategy of the neural network model to reduce the carbon emissions of the neural network model, but also meets the service needs of different users. After the reinforcement learning model is trained, the trained reinforcement learning model can be stored in a preset area. If a user needs to determine the total carbon emissions, the reinforcement learning model used to determine the total carbon emissions can be obtained by accessing the preset area. Step S602: Identify computational graph elements in the neural network model that can execute the parallel strategy. Since the total carbon emissions corresponding to the neural network model are closely related to the computational graph elements in the neural network model that can execute the parallel strategy, to accurately determine the total carbon emissions corresponding to the neural network model, computational graph elements in the neural network model that can execute the parallel strategy can be identified. In some instances, computational graph elements that can execute the parallel strategy can be determined by element type. In this case, identifying computational graph elements in the neural network model that can execute the parallel strategy may include: obtaining the element type of the computational graph element in the computational graph; if the element type is a preset type, determining that the computational graph element can execute the tensor parallel strategy, where the preset type includes at least one of the following: computational graph element multiplication type and computational graph element activation type; if the element type is not the preset type, determining that the computational graph element cannot execute the tensor parallel strategy.Specifically, since there are many types of elements included in the neural network model, not all types of computational graph elements can execute the parallel strategy. Therefore, in order to accurately identify the computational graph elements in the neural network model that can execute the parallel strategy, the element type of the computational graph element corresponding to the neural network model can be first obtained. The element type can be determined by the operating principle of the computational graph element or the preset identifier of the computational graph element. After obtaining the element type, it can be determined whether the element type is a preset type, that is, the element type is analyzed and compared with the preset type, wherein the preset type can include at least one of the following: computational graph element multiplication type, computational graph element activation type. When the element type is the preset type, it is determined that the computational graph element can execute the tensor parallel strategy, that is, when the computational graph element is of the computational graph element multiplication type or the computational graph element activation type, it can be determined that the computational graph element at this time is a computational graph element that can execute the tensor parallel strategy; when the element type is not the preset type, for example: when the element type is an accumulation computational graph element or a bias computational graph element, it can be determined that the computational graph element at this time is not a computational graph element that can execute the tensor parallel strategy. This effectively ensures the accuracy and reliability of determining computation graph elements capable of executing parallel strategies. In other instances, since parallel strategies include not only tensor parallel strategies but also data parallel strategies, computation graph elements capable of executing parallel strategies in a neural network model can be identified based on parallel data requirements. Specifically, in this embodiment, identifying computation graph elements capable of executing parallel strategies in a neural network model can include: obtaining element data of a computation graph element in the computation graph; if the element data corresponds to a parallel data requirement, determining that the computation graph element can execute the data parallel strategy; if the element data does not correspond to a parallel data requirement, determining that the computation graph element cannot execute the data parallel strategy. Among them, in order to accurately identify the computational graph elements in the neural network model that can execute parallel strategies, the element data of the computational graph elements corresponding to the neural network model can be obtained, and then it can be identified whether the element data corresponds to parallel data requirements. In some instances, identifying whether the element data corresponds to parallel data requirements can include: detecting whether a parallel data request corresponding to the element data is obtained; when the parallel data request corresponding to the element data is obtained, it can be determined that the element data corresponds to the parallel data requirement; when the parallel data request corresponding to the element data is not obtained, it can be determined that the element data does not correspond to the parallel data requirement.If it is determined that the element data corresponds to a parallel data requirement, the computation graph element can be determined to be a computation graph element capable of executing a data parallel strategy. If it is determined that the element data does not correspond to a parallel data requirement, the computation graph element can be determined not to be a computation graph element capable of executing a data parallel strategy. This effectively ensures accurate identification of computation graph elements capable of executing a data parallel strategy. Step S603: Determine status information corresponding to the neural network model. The status information includes at least one of the following: remaining system computing power, remaining neural network computing power, computing power requirements, and the amount of data required by the model. When using the neural network model for data processing, the neural network model may correspond to different status information. The aforementioned status information may include at least one of the following: remaining system computing power, remaining neural network computing power, computing power requirements, and the amount of data required by the model. The aforementioned remaining system computing power can be obtained by testing using a preset computing power detection device. The remaining neural network computing power can be determined by combining the total computing power with the occupied computing power and the computation time of the neural network model. Computing power requirements may include intelligent computing power requirements, network computing power requirements, and storage computing power requirements. These requirements may be determined based on user needs and the low-carbon goals to be achieved by the system. The amount of data required by the model may include a collection of model training data, a collection of model inference data, and so on. Step S604: Analyze and process state information and computational graph elements that can execute parallel strategies using a reinforcement learning model to obtain a resource allocation strategy corresponding to the neural network model. The resource allocation strategy may include any one of a parallel strategy and a non-parallel strategy. The parallel strategy may include at least one of the following: a data parallel strategy and a tensor parallel strategy. After obtaining the state information and the computational graph elements that can execute the parallel strategy, the state information and the computational graph elements that can execute the parallel strategy can be analyzed and processed using the reinforcement learning model, so as to obtain a resource allocation strategy corresponding to the neural network model. The resource allocation strategy includes any one of: a parallel strategy and a non-parallel strategy. The parallel strategy includes at least one of the following: a data parallel strategy and a tensor parallel strategy. The above-mentioned data parallel strategy can refer to a strategy for splitting the computational graph element data into multiple sub-data, and then configuring the multiple sub-data to multiple devices for operation and processing; the tensor parallel strategy can refer to a strategy for multiplying the computational graph elements into multiple sub-tensors, and then configuring the multiple sub-tensors to multiple devices for operation and processing. The determined resource allocation strategy can achieve the goal of reducing carbon emissions within a preset time period.It should be noted that when a neural network model operates in different application scenarios, different resource allocation strategies may be obtained. For example, when carbon emissions and electricity costs within a preset time period are low, the resource allocation strategy corresponding to the neural network model is at least one of a tensor parallel strategy and a data parallel strategy. When carbon emissions and electricity costs within the preset time period are high, the resource allocation strategy corresponding to the neural network model includes a non-parallel strategy. This effectively ensures the accuracy and reliability of determining the resource allocation strategy corresponding to the neural network model. Step S605: Determine the total carbon emissions corresponding to the neural network model based on the resource allocation strategy. Because different resource allocation strategies will result in different total carbon emissions for the neural network model, after obtaining the resource allocation strategy, the resource allocation strategy can be analyzed and processed. In some examples, determining the total carbon emissions corresponding to the neural network model based on the resource allocation strategy may include: obtaining a preset mapping relationship for analyzing and processing the resource allocation strategy; and determining the total carbon emissions corresponding to the neural network model based on the preset mapping relationship and the resource allocation strategy. This allows for stable determination of the total carbon emissions corresponding to the neural network model. In some other examples, during data processing by the reinforcement learning model, the state information that the reinforcement learning model needs to process may change. Therefore, during analysis and processing of the state information using the reinforcement learning model, the method in this embodiment may further include: obtaining change information of the state information; and performing a state transition operation based on the change information to obtain a post-transition state. Specifically, during analysis and processing of the state information and computation graph elements using the reinforcement learning model, change information of the state information corresponding to the neural network model may be obtained in real time or periodically. The state change information may be information regarding changes to any state information. For example, the change information may include at least one of the following: first change information corresponding to the remaining computing power of the system, second change information corresponding to the remaining computing power of the neural network, third change information corresponding to the computing power requirement, fourth change information corresponding to the amount of data required by the model, and so on. After acquiring information about changes in state information, state transitions can be performed based on the change information to obtain the post-transformation state. For example, the remaining system computing power, the remaining computational load of the neural network, the computing power requirements, and the data volume required by the model can be obtained. This effectively enables flexible and timely adjustments to state information during data processing by the reinforcement learning model, ensuring its accuracy and further improving the stability and reliability of data processing operations based on the neural network model.In this embodiment, a pre-trained reinforcement learning model for determining resource allocation strategies is obtained, computational graph elements in the neural network model that can execute parallel strategies are identified, and state information corresponding to the neural network model is determined. The reinforcement learning model is then used to analyze and process the state information and computational graph elements to obtain a resource allocation strategy corresponding to the neural network model. Based on the resource allocation strategy, the total carbon emissions corresponding to the neural network model are determined. This effectively ensures the accuracy and reliability of the determination of the total carbon emissions corresponding to the neural network model, further improving the practicality of this method. In specific applications, using a neural network model based on the Pytorch framework as an example, this application embodiment provides a method for determining carbon emissions per unit of intelligent computing power. This method can perform statistical operations on neural network computing power usage based on the computation graph and allocate computing resources based on parallel operations. This method can effectively reduce the carbon emissions of neural network models. Specifically, the method for determining the carbon emissions per unit of intelligent computing power mainly includes the following two parts: (1) statistical operations on the computational amount of the neural network model based on the computational graph, which can analyze the detailed computational amount data in each process of the neural network model; (2) computing resource allocation based on parallel strategies, making full use of the flexibility of the neural network model to achieve the low-carbon goal. The first part mentioned above can include the following steps: Step 11: Obtain the neural network model corresponding to the cloud computing task. Step 12: Decompose the neural network model into a computational graph composed of computational graph elements. Taking a Long Short-Term Memory (LSTM) network model as an example of a neural network model, the computational graph decomposed from the neural network model can be shown in Figure 7. The computational graph may include multiple basic computational graph elements, such as bias computational graph elements, accumulation computational graph elements, multiplication or addition computational graph elements, and transposition computational graph elements. These basic computational graph elements may be nodes or edges used to construct the computational graph. Nodes may include at least one of the following: transfer functions, accumulation operations, multiplication operations, bias operations, and transposition operations. Edges may include at least one of the following: model parameters, input matrices, and the like. By decomposing the neural network model into basic computational graph elements, the computational graph has good scalability and can be applied to a variety of different neural network models. Step 13: Identify parallel computational graph elements included in the computational graph.Since parallel computation graph elements refer to nodes or edges that can be operated in parallel, to accurately identify parallel computation graph elements in a neural network model, it is possible to determine whether any two computation graph elements in the neural network model have a data dependency relationship, or to determine whether any two computation graph elements in the neural network model belong to the same data layer. Specifically, to accurately identify whether any two computation graph elements in the neural network model have a data dependency relationship, the computation graph elements in the computation graph can be ID-ized and a corresponding storage matrix generated for them:

[0005] 0 1 ... 1'

[0006] M = ° ° ■" 1 £ R nxn

[0007] 0 0 ... 0

[0008] .0 0 ... 1. The horizontal and vertical elements of the matrix M represent n computational graph elements in the computational graph. Each element in the matrix indicates whether a data dependency exists between the computational graph elements. For example, if an element is "1," it indicates that a data dependency exists between the two computational graph elements; if an element is "0," it indicates that no data dependency exists. After determining whether a data dependency exists between any two computational graph elements, a directed acyclic graph corresponding to the computational graph can be obtained based on the data dependency. This directed acyclic graph can then be used to determine whether any two computational graph elements are in the same data layer. If any two computational graph elements are in the same data layer and there is no data dependency between them, they can be determined to be parallel computational graph elements that can be processed in parallel. As shown in Figure 7, in the computational graph corresponding to the LSTM model, two multiplication or addition computational graph elements can be determined to be parallel computational graph elements that can be computed in parallel. If any two computational graph elements are located in different data layers, or if there is a data dependency between the two computational graph elements, these two computational graph elements can be determined as non-parallel computational graph elements that cannot be processed in parallel. This effectively identifies the parallel computational graph elements included in the computational graph and determines the number of parallel computational graph elements included in the computational graph. Step 14: Based on the computational graph, obtain the raw computational load corresponding to the neural network model during operation. After obtaining the computational graph, traverse the computational graph to obtain the total computational load required for a forward propagation and backward propagation operation of the neural network model, thereby obtaining the computational cost of each process in the neural network model. Forward propagation refers to the process of inputting input data into the neural network model and obtaining the output of the neural network model; backward propagation refers to the process of calculating the gradients of various parameters based on the output of the neural network model after obtaining the output. Specifically, for the nodes and edges included in the computation graph, different nodes and edges correspond to different computational amounts. For example, the computational amount corresponding to the multiplication of computation graph elements is the first computational amount, and the computational amount corresponding to the activation function is the second computational amount. Since the computational amount corresponding to the multiplication operation of computation graph elements and the activation function operation is large, the impact of the above-mentioned multiplication operation of computation graph elements and the activation function operation on the total computational amount is ignored. The computational amount corresponding to other operations other than the multiplication operation of computation graph elements and the activation function operation is small. Therefore, the impact of the above-mentioned non-multiplication operation of computation graph elements and the non-activation function operation on the total computational amount can be ignored.Note that for a neural network model, the above-mentioned multiplication operation of computational graph elements and activation function operation are involved in both the forward propagation process and the gradient calculation process of the backward propagation. The computational graph elements corresponding to the model parameters are involved in the update process during the backward propagation. Among them, the computational amounts of the forward propagation process and the backward propagation process can be respectively expressed as: where, qi represents the i-th calculation... is the multiplication operation of computational graph elements; when fj = 0, the data item corresponding to gj does not exist. gj represents the computational graph element corresponding to the model parameter of the i-th computational graph, and qi represents the i-th calculation... k represents the matrix parameter corresponding to the multiplication operation of computational graph elements. For example, when there are multiplication operations of matrix m * k and matrix k * n in the neural network model, the matrix parameter corresponding to the matrix multiplication operation can be determined. vi is the number of operations corresponding to the input computational graph element, wi is the number of operations required for the computational graph element corresponding to the model parameter, and u is the number of operations corresponding to the transfer relationship. Through the above formula, the computational amount A of the neural network model after one forward propagation process and the computational amount B of the neural network model after one backward propagation process can be stably obtained. It should be noted that during the backward propagation process, the computational amount required for the multiplication operation of computational graph elements is twice that of the forward propagation process because gradients need to be calculated for both the input computational graph elements and the model parameters at the same time. In addition, the computational amount generated during the parameter update process needs to be calculated additionally, which is related to the type of optimizer. The original computational amount of the neural network model running can be expressed as the training and inference processes, and can be specifically obtained through the following formula:

[0009] S = infer x A + train x (A + B) where, S is the original computational amount, infer is the number of inference stages experienced by the neural network model, train is the number of training stages experienced by the neural network model, A is the computational amount of the neural network model after one forward propagation process, and B is the computational amount of the neural network model after one backward propagation process. Step 15: Based on the parallel computational graph elements and the original computational amount included in the computational graph, determine the correction coefficient for correcting the original computational amount. Specifically, the correction coefficient for correcting the original computational amount can be obtained through the following formula: a*(S) = (W * S2) * where, <ZK(S) is the correction coefficient for correcting the original computational amount, SK is the computational amount corresponding to the non-computational graph element, S is the original computational amount corresponding to the neural network model, S - SR is the computational amount corresponding to the parallel computational graph element, p Kis the preset coefficient corresponding to the parallel computation graph element, which can be determined by regression. K is a pre-configured learnable parameter, which can be determined by regression. The above formula indicates that the parallelizable computation amount can be removed from the total original computation amount, which can improve the accuracy and reliability of determining the total computation amount corresponding to the neural network model during operation. Step 16: Based on the correction coefficient used to correct the original computation amount and the original computation amount, determine the total computation amount corresponding to the neural network model during operation. Specifically, the total computation amount corresponding to the neural network model during operation can be obtained by the following formula: Q = a K(S) x S, where the product of the original computational effort and the correction coefficient is determined as the total computational effort corresponding to the neural network model during operation. After determining the total computational effort corresponding to the neural network model during operation, to fully utilize the temporal and spatial flexibility of the neural network model to achieve low-carbon goals, the computing resource allocation strategy can be adjusted so that the neural network model maximizes utilization of idle resources during periods of low electricity and carbon prices, while minimizing computing resource consumption during periods of low electricity and carbon prices. Furthermore, to ensure the quality of service for data processing operations, certain requirements may still be placed on the neural network's runtime. To achieve the above effects, this application embodiment can deploy the neural network model on different devices at different times to allocate computing resources. Because different parallelization strategies generate different additional computing power costs, the parallelization strategy can be determined based on a pre-trained neural network model. Specifically, the method may include the following steps: Step 21: Constructing a Deep Reinforcement Learning (DRL) model for determining the resource allocation strategy. To simplify DRL model construction, we can assume that the multiple devices running the neural network model are of the same type and that the computing power curves of each device remain consistent. Model training can then be performed using each neural network model's resource allocation strategy as the decision variable, with the optimization objective of reducing total carbon emissions, electricity costs, and the combined loss of service quality within a preset time period. This will yield a DRL model for determining resource allocation strategies. In general, the basic elements of DRL model construction include: a state space, an optimization objective, and state transitions. The state space represents the state information corresponding to the neural network model's data processing operations and is used to determine its computing resource allocation strategy. In this application embodiment, state information can be determined based on the load of the system center or platform center and the parsed neural network data. Specifically, this state information may include: the remaining computing power of the data center (i.e., the utilization rate of each computing resource in the data center (graphics processing unit GPU, parallel processing unit (TPU)), etc.), the remaining computational load of the neural network (which can be obtained by combining the total computational load and the unfinished computational load), the network computing power requirement (which can be obtained by analyzing the computation graph), the storage computing power requirement, and the amount of training or inference data. In some examples, the state space can be represented as: Furthermore, the corresponding curve mapping relationships for different application scenarios, different neural network models, and different data processing operations effectively ensure the accurate and reliable determination of the remaining computing power of the data center and the remaining computing capacity of the neural network. The action space represents the actions that each neural network model can perform. In this embodiment, it represents the parallel solutions that can be selected by the neural network model and can be expressed as:

[0010] A — {, data, tensor] where 4 represents the action space, indicating that the neural network model does not adopt any parallel strategy, and data is used to indicate that the neural network model selects a data parallel strategy. When data = n, it can be determined that the data parallel strategy selected by the neural network model requires data to be distributed across n devices; when tensor = m, it can be determined that the tensor parallel strategy selected by the neural network model allocates data to n devices. State transitions can represent the transitions in the neural network model's state at each moment. In this embodiment, they represent changes in the data center's occupancy and the amount of neural network computation. For example, during data processing using the neural network model, the data center's computing resources and idleness can change in real time. The objective function can be used to guide parameter updates in the DRL model. To minimize energy costs and carbon emissions, the optimization objective can be the combined total carbon emissions, electricity consumption costs, and service quality losses within a preset time period. Step 22: Identify computational graph elements in the neural network model that can execute parallel strategies. For neural network models, common parallel strategies currently include data parallelism and tensor parallelism. Data parallelism refers to distributing a copy of the model and a portion of the data used for training or inference to multiple devices for parallel computation. Since data parallelism does not change the model structure, it can be adopted by most models. Tensor parallelism, on the other hand, refers to splitting and distributing the computational graph element operations across multiple devices. Therefore, it can only be adopted by elements within the neural network that involve tensor operations (e.g., Addmm operations). Based on these parallel strategies, computation graph elements within the neural network model that can execute parallelism can be identified. In some instances, the element type of the computation graph element corresponding to the neural network model can be first obtained. If the element type is a preset type, the computation graph element is determined to be a computation graph element that can execute tensor parallelism. The preset type includes at least one of the following: computation graph element multiplication type and computation graph element activation type. If the element type is not a preset type, the computation graph element is determined not to be a computation graph element that can execute tensor parallelism. In other instances, the element data of the computational graph element corresponding to the neural network model can be first obtained; when the element data corresponds to a parallel data requirement, the computational graph element is determined to be a computational graph element capable of executing a data parallel strategy; when the computational graph element data does not correspond to a parallel data requirement, the computational graph element is determined not to be a computational graph element capable of executing a data parallel strategy. Through the above operations, the computational graph elements in the neural network model that can execute a parallel strategy can be accurately identified.Step 23: Determine state information corresponding to the neural network model. The state information includes at least one of the following: remaining system computing power, remaining neural network computing power, computing power requirements (including network computing power requirements and storage computing power requirements), and the amount of data required by the model. Step 24: Analyze and process the state information and computation graph elements using a reinforcement learning model to obtain a resource allocation strategy corresponding to the neural network model. The resource allocation strategy includes any one of a parallel strategy and a non-parallel strategy. The parallel strategy includes at least one of a data parallel strategy and a tensor parallel strategy. Step 25: Determine the total carbon emissions corresponding to the neural network model based on the resource allocation strategy. Determining the total carbon emissions corresponding to the neural network model based on the resource allocation strategy may include: obtaining a preset mapping relationship for analyzing and processing the resource allocation strategy; and determining the total carbon emissions corresponding to the neural network model based on the preset mapping relationship and the resource allocation strategy. Step 30: After obtaining the total carbon emissions and total computing power, the carbon emissions per unit computing power corresponding to the neural network model may be determined based on the total carbon emissions and total computing power. Specifically, the carbon emissions per unit computing power can be obtained by R = 3, where R is the carbon emissions per unit computing power, C is the total carbon emissions, and Q is the total computing power. This allows for the stable acquisition of the carbon emissions per unit computing power corresponding to the neural network model.The technical solution provided by this application embodiment can accurately measure and optimize carbon emissions per unit computing power at the topological level of the intelligent computing model. Specifically, it can perform neural network computing power usage statistics based on the computational graph, decomposing the topological structure of the neural network model into a computational graph composed of basic computational graph elements. The intelligent computing power and storage computing power required at the same time are counted based on internal data dependencies, thereby accurately counting the computational load of the neural network model. The statistical accuracy is verified by testing the theoretical and actual running times of the neural network model on a certain device, with an error of no more than 10%. After obtaining the computational load, the potential impact of various computing resource allocation schemes on data center carbon emissions is considered, and the spatiotemporal flexibility of the neural network model is fully utilized to achieve the goal of reducing carbon emissions. Specifically, based on the computational graph and computational load of the neural network, the additional network computing power cost and the reduced intelligent computing power and storage computing power cost generated by strategies such as data parallelism and tensor parallelism are calculated. The remaining computational load of the neural network model is comprehensively considered, and the DRL method is used to determine the optimal parallel processing strategy for each neural network model. For example, different parallel processing strategies can be used to allow the neural network model to suspend operation to a certain extent when carbon prices and electricity prices are high. It can also be dispatched to other data centers with lower carbon emission factors (requiring a preset network computing capacity). This helps reduce the carbon emissions and electricity costs generated by the operation of user-submitted neural network models over a period of time, thereby achieving the effect of reducing overall carbon emissions and electricity costs, further improving the practicality of the method and facilitating market promotion and application. Figure 8 is a flow chart of a method for determining the total computing capacity of a cloud computing task, provided in an embodiment of the present disclosure. Referring to Figure 8 , this embodiment provides a method for determining the total computing capacity of a cloud computing task. The method may be performed by a device for determining the total computing capacity of a cloud computing task. It is understood that the device for determining the total computing capacity of a cloud computing task may be implemented as software or a combination of software and hardware. Specifically, when the device for determining the total computing capacity of a cloud computing task is implemented as hardware, it may be various electronic devices capable of performing the operation of determining the total computing capacity, including but not limited to tablet computers, personal computers (PCs), servers, and the like. When the device for determining the total computing capacity of a cloud computing task is implemented as software, it may be installed in the electronic devices listed above. Based on the above-mentioned device for determining the total computing power of a cloud computing task, the method for determining the total computing power of a cloud computing task in this embodiment may include the following steps: Step S801: Obtain a neural network model corresponding to the cloud computing task.Step S802: Determine a computation graph corresponding to the neural network model. The computation graph is composed of multiple computation graph elements, at least some of which can be operated in parallel. Each computation graph element is any one of the following: a node, an edge between nodes. Step S803: Determine the total computational effort corresponding to the neural network model during operation based on the parallel computation graph elements included in the computation graph. The implementation principles and effects of the method in this embodiment are similar to the specific implementation principles and effects of the method steps in the embodiment shown in FIG2 . For details, please refer to the above description and will not be repeated here. It should be noted that the method in this embodiment may also include methods that can execute the embodiments shown in FIG2 - FIG7 . For portions not described in detail in this embodiment, please refer to the relevant description of the embodiments shown in FIG2 - FIG7 . The implementation process and technical effects of this technical solution are described in the embodiments shown in FIG2 - FIG7 , and will not be repeated here. The method for determining the total computing power corresponding to the cloud computing task in this embodiment obtains the neural network model corresponding to the cloud computing task; determines the computation graph corresponding to the neural network model, and determines the total computing power corresponding to the neural network model during operation based on the parallel computation graph elements included in the computation graph; this effectively implements the statistical operation of the neural network computing power occupancy based on the computation graph. Since the impact of the parallel computation graph elements on the total computing power is taken into account, the accuracy and reliability of the determination of the total computing power is effectively guaranteed; then, based on the obtained total computing power, the carbon emissions per unit computing power corresponding to the neural network model can be determined, thereby effectively improving the practicality of the method. 9 is a schematic structural diagram of a device for determining carbon emissions per unit computing power provided by an embodiment of the present disclosure; with reference to FIG9 , this embodiment provides a device for determining carbon emissions per unit computing power, which is used to execute the method for determining carbon emissions per unit computing power shown in FIG2 . The determining device may include: a first acquisition component 11 for acquiring a neural network model corresponding to a cloud computing task; a first determination component 12 for determining a computational graph corresponding to the neural network model, where the computational graph is composed of multiple computational graph elements, at least some of which can operate in parallel, and each computational graph element is any one of the following: a node, an edge between nodes; the first determination component 12 is further used to determine the total computational amount corresponding to the neural network model during operation based on the parallel computational graph elements included in the computational graph; the first determination component 12 is further used to determine the total carbon emissions corresponding to the neural network model; and a first processing component 13 is used to determine the carbon emissions per unit computing power corresponding to the neural network model based on the total carbon emissions and the total computational amount.In some instances, after determining the computation graph corresponding to the neural network model, the first processing component 13 in this embodiment is configured to perform the following steps: based on the computation graph, determine the computation layer to which each computation graph element belongs and whether there is a data dependency between any two computation graph elements; if two computation graph elements belong to the same computation layer and there is no data dependency between the two computation graph elements, determine that the two computation graph elements are parallel computation graph elements. In some instances, when the first determination component 12 determines the total computation amount corresponding to the neural network model during operation based on the parallel computation graph elements included in the computation graph, the first determination component 12 is configured to: obtain the raw computation amount corresponding to the neural network model during operation based on the computation graph; and modify the raw computation amount using the parallel computation graph elements included in the computation graph to obtain the total computation amount corresponding to the neural network model during operation. In some instances, when the first determination component 12 obtains the raw computational load corresponding to the neural network model during operation based on the computation graph, the first determination component 12 is configured to: traverse the computation graph to obtain a first computational load corresponding to one forward propagation operation and a second computational load corresponding to one backward propagation operation of the neural network model; determine the number of training stages and the number of inference stages corresponding to the neural network model; and determine the raw computational load corresponding to the neural network model during operation based on the number of training stages, the number of inference stages, the first computational load, and the second computational load. In some instances, when the first determination component 12 determines the raw computational load corresponding to the neural network model during operation based on the number of training stages, the number of inference stages, the first computational load, and the second computational load, the first determination component 12 is configured to: obtain a computational load sum value between the first computational load and the second computational load; determine a first product value between the computational load sum value and the number of training stages, and a second product value between the first computational load and the number of inference stages; and determine the sum of the first product value and the second product value as the raw computational load corresponding to the neural network model during operation.In some instances, when the first determination component 12 traverses the computation graph to obtain a first computational amount corresponding to a forward propagation operation of the neural network model, the first determination component 12 is used to perform: traversing the computation graph to obtain the input computation graph elements, computation graph elements corresponding to the model parameters, matrix parameters corresponding to the matrix multiplication operation, the number of matrix multiplication operations, and the number of operations of the computation graph elements corresponding to the model parameters, which are included in a forward propagation operation of the neural network model; and determining the first computational amount corresponding to a forward propagation operation of the neural network model based on the input computation graph elements, the computation graph elements corresponding to the model parameters, the matrix parameters corresponding to the matrix multiplication operation, the number of matrix multiplication operations, and the number of operations of the computation graph elements corresponding to the model parameters. In some instances, when the first determination component 12 traverses the computation graph to obtain the second computational amount corresponding to a back propagation operation of the neural network model, the first determination component 12 is configured to: traverse the computation graph to obtain the input computation graph elements, computation graph elements corresponding to model parameters, transfer relationships, matrix parameters corresponding to matrix multiplication operations, the number of matrix multiplication operations, the first number of operations of the computation graph elements corresponding to the model parameters, and the second number of operations corresponding to the transfer relationships, which are included in a back propagation operation of the neural network model; and determine the second computational amount corresponding to a back propagation operation of the neural network model based on the input computation graph elements, computation graph elements corresponding to model parameters, transfer relationships, matrix parameters corresponding to matrix multiplication operations, the number of matrix multiplication operations, the first number of operations of the computation graph elements corresponding to the model parameters, and the second number of operations corresponding to the transfer relationships. In some instances, when the first determination component 12 uses the parallel calculation graph elements included in the calculation graph to correct the original calculation amount and obtain the total calculation amount corresponding to the neural network model during the operation process, the first determination component 12 is used to perform: based on the parallel calculation graph elements included in the calculation graph and the original calculation amount, determine the correction coefficient used to correct the original calculation amount; based on the correction coefficient and the original calculation amount, obtain the total calculation amount corresponding to the neural network model during the operation process.In some instances, when the first determination component 12 determines a correction coefficient for correcting the original computational amount based on the parallel computational graph elements and the original computational amount included in the computational graph, the first determination component 12 is configured to perform the following steps: obtaining the parallel computational amount corresponding to the parallel computational graph elements in the computational graph and the weight coefficient corresponding to the parallel computational amount; determining the non-parallel computational amount corresponding to the neural network model based on the parallel computational amount and the original computational amount; and determining the correction coefficient for correcting the original computational amount based on the non-parallel computational amount, the parallel computational amount, the weight coefficient, and the original computational amount. In some instances, when the first determination component 12 determines a correction coefficient for correcting the original computation amount based on the non-parallel computation amount, the parallel computation amount, the weight coefficient, and the original computation amount, the first determination component 12 is configured to: obtain a learnable parameter for determining the correction coefficient; determine a ratio between the product of the parallel computation amount and the weight coefficient, the product of the product and the sum of the non-parallel computation amount, and the sum of the computation amount and the original computation amount; and determine the correction coefficient for correcting the original computation amount using the ratio as a base and the learnable parameter as an exponent. In some instances, when the first determination component 12 determines the total carbon emissions corresponding to the neural network model, the first determination component 12 is used to execute: obtaining a pre-trained reinforcement learning model for determining a resource allocation strategy, where the reinforcement learning model is trained using the resource allocation strategy of the neural network model as a decision variable and reducing carbon emissions within a preset time period as an optimization goal; identifying computational graph elements in the neural network model that can execute parallel strategies; determining state information corresponding to the neural network model, where the state information includes at least one of the following: remaining computing power of the system, remaining computing power of the neural network, computing power requirements, and the amount of data required by the model; using the reinforcement learning model to analyze and process the state information and computational graph elements to obtain a resource allocation strategy corresponding to the neural network model, where the resource allocation strategy includes any one of: a parallel strategy and a non-parallel strategy, and the parallel strategy includes at least one of: a data parallel strategy and a tensor parallel strategy; and determining the total carbon emissions corresponding to the neural network model based on the resource allocation strategy. In some instances, when the first determination component 12 identifies a computational graph element in a neural network model that can execute a parallel strategy, the first determination component 12 is used to execute: obtaining an element type of the computational graph element in the computational graph; when the element type is a preset type, determining that the computational graph element can execute a tensor parallel strategy, wherein the preset type includes at least one of the following: a computational graph element multiplication type, a computational graph element activation type; when the element type is not a preset type, determining that the computational graph element cannot execute a tensor parallel strategy.In some instances, when the first determination component 12 identifies a computational graph element in a neural network model that can execute a parallel strategy, the first determination component 12 is configured to: obtain element data for the computational graph element; if the element data corresponds to a parallel data requirement, determine that the computational graph element can execute a data parallel strategy; if the element data does not correspond to a parallel data requirement, determine that the computational graph element cannot execute a data parallel strategy. In some instances, the first determination component 12 is configured to: obtain state information change information; and perform a state transition operation based on the change information to obtain a post-transition state. The apparatus shown in FIG9 can execute the method of the embodiments shown in FIG1-FIG7 . For portions not described in detail in this embodiment, reference is made to the relevant description of the embodiments shown in FIG1-FIG7 . The execution process and technical effects of this technical solution are described in the embodiments shown in FIG1-FIG7 and will not be further elaborated here. In one possible design, the structure of the apparatus for determining carbon emissions per unit computing power of a cloud computing task shown in FIG9 can be implemented as an electronic device, which can be a controller, a personal computer, a partition, or other devices. As shown in FIG10 , the electronic device can include a first processor 21 and a first memory 22. The first memory 22 is used to store a program corresponding to the method for determining the carbon emissions per unit of computing power of an electronic device performing a cloud computing task provided in the embodiments shown in Figures 1-7 . The first processor 21 is configured to execute the program stored in the first memory 22. The program includes one or more computer instructions. When executed by the first processor 21, the one or more computer instructions can implement the following steps: obtaining a neural network model corresponding to the cloud computing task; determining a computational graph corresponding to the neural network model, the computational graph consisting of multiple computational graph elements, at least some of which can be operated in parallel, each computational graph element being any one of the following: a node, an edge between nodes; determining the total computational load corresponding to the neural network model during operation based on the parallel computational graph elements included in the computational graph; determining the total carbon emissions corresponding to the neural network model; and determining the carbon emissions per unit of computing power corresponding to the neural network model based on the total carbon emissions and the total computational load. Furthermore, the first processor 21 is further configured to execute all or part of the steps in the embodiments shown in Figures 1-7 . The electronic device may also include a first communication interface 23 for communicating with other devices or a communication network. In addition, an embodiment of the present disclosure provides a computer storage medium for storing computer software instructions used by an electronic device, which includes a program for executing the method for determining carbon emissions per unit computing power of a cloud computing task in the embodiments shown in Figures 1 to 7 above.Furthermore, an embodiment of the present disclosure provides a computer program product comprising: a computer-readable storage medium storing computer instructions, which, when executed by one or more processors, causes the one or more processors to perform the steps of the method for determining carbon emissions per unit computing power of a cloud computing task in the method embodiments shown in Figures 1-7 . Figure 11 is a schematic structural diagram of a device for determining the total computing power of a cloud computing task provided by an embodiment of the present disclosure. Referring to Figure 11 , this embodiment provides a device for determining the total computing power of a cloud computing task, which is configured to perform the method for determining the total computing power of a cloud computing task shown in Figure 8 . The device may include: a second acquisition component 31 for acquiring a neural network model corresponding to the cloud computing task; a second determination component 32 for determining a computation graph corresponding to the neural network model, the computation graph comprising multiple computation graph elements, at least some of which can operate in parallel, each computation graph element being any one of the following: a node or an edge between nodes; and a second processing component 33 for determining the total computation power corresponding to the neural network model during execution based on the parallel computation graph elements included in the computation graph. The device shown in FIG11 can execute the method of the embodiment shown in FIG8 . For portions not described in detail in this embodiment, please refer to the relevant description of the embodiment shown in FIG8 . The execution process and technical effects of this technical solution are described in the embodiment shown in FIG8 and will not be repeated here. In one possible design, the structure of the device for determining the total computing power of a cloud computing task shown in FIG11 can be implemented as an electronic device, which can be a controller, a personal computer, a partition, or other devices. As shown in FIG12 , the electronic device may include a second processor 41 and a second memory 42. The second memory 42 is used to store a program corresponding to the electronic device executing the method for determining the total computing power of a cloud computing task provided in the embodiment shown in FIG8 . The second processor 41 is configured to execute the program stored in the second memory 42. The program includes one or more computer instructions, wherein when executed by the second processor 41, the one or more computer instructions can implement the following steps: obtaining a neural network model corresponding to the cloud computing task; determining a computation graph corresponding to the neural network model, wherein the computation graph is composed of multiple computation graph elements, at least some of the multiple computation graph elements can be operated in parallel, and each computation graph element is any one of the following: a node, an edge between nodes; and determining a total computational load corresponding to the execution of the neural network model based on the parallel computation graph elements included in the computation graph. Furthermore, the second processor 41 is further configured to execute all or part of the steps in the embodiment shown in FIG. 8 .The electronic device may also include a second communication interface 43 for communicating with other devices or a communication network. Furthermore, embodiments of the present disclosure provide a computer storage medium for storing computer software instructions used by the electronic device, including a program for executing the method for determining the total computing power of a cloud computing task in the embodiment shown in FIG. Furthermore, embodiments of the present disclosure provide a computer program product comprising: a computer-readable storage medium storing computer instructions. When the computer instructions are executed by one or more processors, the one or more processors execute the steps of the method for determining the total computing power of a cloud computing task in the embodiment shown in FIG. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, storage, and display, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. The collection, use, and processing of the relevant data must comply with the relevant laws, regulations, and standards of relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or reject. The device embodiments described above are merely illustrative. Components described as separate parts may or may not be physically separate, and components shown as components may or may not be physical components, i.e., they may be located in one location or distributed across multiple network components. Some or all of these components may be selected based on actual needs to achieve the objectives of the present embodiments. Persons of ordinary skill in the art will be able to understand and implement the present embodiments without inventive effort. Through the above descriptions of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using the necessary general-purpose hardware platform, or through a combination of hardware and software. Based on this understanding, the essence of the above technical solutions, or the portion that contributes to the prior art, can be embodied in the form of a computer product. The present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions.These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable device to produce a machine, such that the instructions executed by the processor of the computer or other programmable device produce means for implementing the functions specified in one or more flows of the flowchart and / or one or more blocks of the block diagram. These computer program instructions can also be stored in a computer-readable memory capable of directing the computer or other programmable device to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means, which implement the functions specified in one or more flows of the flowchart and / or one or more blocks of the block diagram. These computer program instructions can also be loaded onto a computer or other programmable device, such that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, such that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows of the flowchart and / or one or more blocks of the block diagram. In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-permanent storage in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media. Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can be implemented using any method or technology for information storage. Information can be computer-readable instructions, data structures, program components, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device. As defined in this article, computer-readable media does not include transient computer-readable media (transit media), such as modulated data signals and carrier waves.Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present disclosure and are not intended to limit the present disclosure. Although the present disclosure has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they may modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. However, such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present disclosure. Industrial Applicability: The solution provided by the embodiments of the present disclosure can be applied to the process of determining and reducing carbon emissions in a data processing system or data processing platform. This involves obtaining a neural network model corresponding to a cloud computing task, determining a computational graph corresponding to the neural network model, and determining the total computational load corresponding to the neural network model during operation based on the parallel computational graph elements included in the computational graph. This effectively implements a computational graph-based statistical operation of neural network computing power usage. By taking into account the impact of the parallel computational graph elements on the total computational load, the accuracy and reliability of the determination of the total computational load are effectively guaranteed. The total carbon emissions corresponding to the neural network model are then determined, and the unit computing power carbon emissions corresponding to the neural network model are determined based on the total carbon emissions and the total computational load, thereby effectively improving the accuracy and reliability of the determination of unit computing power carbon emissions. The obtained unit computing power carbon emissions can be used as an indicator for uniformly measuring computing power, energy consumption, and carbon emissions in intelligent computing scenarios. Once the unit computing power carbon emissions are obtained, the unit computing power carbon emissions can be used to guide the carbon emissions analysis of the neural network model. This not only ensures the flexibility and reliability of the neural network model operation but also helps reduce the carbon emissions of the system or platform. It meets the needs of green environmental protection and further improves the practicality of the method.

Claims

26 Claims 1. A method for determining carbon emissions per unit computing power of a cloud computing task, comprising: Obtain the neural network model corresponding to the cloud computing task; Determine a computational graph corresponding to the neural network model, wherein the computational graph is composed of a plurality of computational graph elements, at least a portion of the plurality of computational graph elements can operate in parallel, and each computational graph element is any one of the following: a node, an edge between nodes; based on the parallel computational graph elements included in the computational graph, determine a total computational amount corresponding to the neural network model during operation; determine a total carbon emission corresponding to the neural network model; and based on the total carbon emission and the total computational amount, determine a unit computing power carbon emission corresponding to the neural network model.

2. The method according to claim 1, wherein: After determining the computation graph corresponding to the neural network model, the method further includes: based on the computation graph, determining the computation layer to which each computation graph element belongs, and whether there is a data dependency relationship between any two computation graph elements; in response to the two computation graph elements belonging to the same computation layer and the absence of the data dependency relationship between the two computation graph elements, determining that the two computation graph elements are parallel computation graph elements.

3. The method according to claim 1, wherein: Determining a total computational amount corresponding to the neural network model during operation based on the parallel computational graph elements included in the computational graph includes: obtaining an original computational amount corresponding to the neural network model during operation based on the computational graph; and correcting the original computational amount using the parallel computational graph elements included in the computational graph to obtain the total computational amount.

4. The method according to claim 3, wherein: Based on the computation graph, obtaining an original computational amount corresponding to the neural network model during operation includes: traversing the computation graph to obtain a first computational amount corresponding to a forward propagation operation of the neural network model and a second computational amount corresponding to a backpropagation operation of the neural network model; determining the number of training stages and the number of inference stages corresponding to the neural network model; and determining the original computational amount corresponding to the neural network model during operation based on the number of training stages, the number of inference stages, the first computational amount, and the second computational amount.

5. The method according to claim 4, wherein: Based on the number of training stages, the number of inference stages, the first calculation amount and the second calculation amount, determining the original calculation amount corresponding to the neural network model during operation, including: obtaining a calculation amount sum value between the first calculation amount and the second calculation amount; determining a first product value between the calculation amount sum value and the number of training stages, and a second product value between the first calculation amount and the number of inference stages; and determining the sum value between the first product value and the second product value as the original calculation amount corresponding to the neural network model during operation.

6. The method according to claim 4, wherein: Performing a traversal operation on the computation graph to obtain a first computational amount corresponding to one forward propagation operation of the neural network model includes: Performing a traversal operation on the computation graph to obtain input computation graph elements included in one forward propagation operation of the neural network model, computation graph elements corresponding to model parameters of the neural network model, matrix parameters corresponding to the matrix multiplication operation, the number of matrix multiplication operations, and the number of operations of the computation graph elements corresponding to the model parameters; Based on the input calculation graph elements, the calculation graph elements corresponding to the model parameters, the matrix parameters corresponding to the matrix multiplication operations, the number of matrix multiplication operations and the number of operations of the calculation graph elements corresponding to the model parameters, the first calculation amount corresponding to the neural network model after one forward propagation operation is determined.

7. The method according to claim 4, wherein: Performing a traversal operation on the computation graph to obtain a second computational amount corresponding to a back-propagation operation of the neural network model, including: performing a traversal operation on the computation graph to obtain the input computation graph elements, computation graph elements corresponding to model parameters of the neural network model, transfer relationships, matrix parameters corresponding to matrix multiplication operations, the number of matrix multiplication operations, the first number of operations of the computation graph elements corresponding to the model parameters, and the second number of operations corresponding to the transfer relationship, included in the back-propagation operation of the neural network model; determining the second computational amount corresponding to a back-propagation operation of the neural network model based on the input computation graph elements, the computation graph elements corresponding to the model parameters, the transfer relationships, the matrix parameters corresponding to the matrix multiplication operations, the number of matrix multiplication operations, the first number of operations of the computation graph elements corresponding to the model parameters, and the second number of operations corresponding to the transfer relationships.

8. The method according to claim 3, wherein: Utilizing the parallel computation graph elements included in the computation graph, the original computation amount is corrected to obtain the total computation amount, including: determining a correction coefficient for correcting the original computation amount based on the parallel computation graph elements included in the computation graph and the original computation amount; and obtaining the total computation amount corresponding to the neural network model during operation based on the correction coefficient and the original computation amount.

9. The method according to claim 8, wherein: Based on the parallel calculation graph elements and the original calculation amount included in the calculation graph, determining a correction coefficient for correcting the original calculation amount, including: obtaining the parallel calculation amount corresponding to the parallel calculation graph elements in the calculation graph, and the weight coefficient corresponding to the parallel calculation amount; based on the parallel calculation amount and the original calculation amount, determining the non-parallel calculation amount corresponding to the neural network model; based on the non-parallel calculation amount, the parallel calculation amount, the weight coefficient and the original calculation amount, determining the correction coefficient for correcting the original calculation amount.

10. The method according to claim 9, wherein: Determining the correction coefficient for correcting the original calculation amount based on the non-parallel calculation amount, the parallel calculation amount, the weight coefficient and the original calculation amount includes: obtaining a learnable parameter for determining the correction coefficient; determining the product value of the parallel calculation amount and the weight coefficient, the ratio between the product value and the sum of the calculation amount of the non-parallel calculation amount, and the ratio between the sum of the calculation amount and the original calculation amount; using the ratio as a base and the learnable parameter as the exponential base of the exponent to determine the correction coefficient for correcting the original calculation amount.

11. The method according to any one of claims 1 to 0, wherein: Determining the total carbon emissions corresponding to the neural network model includes: obtaining a pre-trained reinforcement learning model for determining a resource allocation strategy, wherein the reinforcement learning model is trained using the resource allocation strategy of the neural network model as a decision variable and reducing carbon emissions within a preset time period as an optimization objective; Identify the computational graph elements in the neural network model that can execute parallel strategies; determine state information corresponding to the neural network model, wherein the state information includes at least one of the following: remaining computing power of the system, remaining computing power of the neural network, computing power requirements, and the amount of data required by the model; use the reinforcement learning model to analyze and process the state information and the computational graph elements to obtain a resource allocation strategy corresponding to the neural network model, wherein the resource allocation strategy includes any one of: a parallel strategy and a non-parallel strategy, and the parallel strategy includes at least one of the following: a data parallel strategy and a tensor parallel strategy; based on the resource allocation strategy, determine the total carbon emissions corresponding to the neural network model.

12. The method according to claim 11, wherein: Identifying the computational graph element in the neural network model that can execute the parallel strategy includes: obtaining the element type of the computational graph element in the computational graph; in response to the element type being a preset type, determining that the computational graph element can execute the tensor parallel strategy, wherein the preset type includes at least one of the following: a computational graph element multiplication type, a computational graph element activation type; in response to the element type not being a preset type, determining that the computational graph element cannot execute the tensor parallel strategy.

13. The method according to claim 11, wherein: Identifying the computational graph element in the neural network model that can execute the parallel strategy includes: obtaining element data of the computational graph element in the computational graph; in response to the element data corresponding to a parallel data requirement, determining that the computational graph element can execute the data parallel strategy; in response to the element data not corresponding to the parallel data requirement, determining that the computational graph element cannot execute the data parallel strategy.

14. A method for determining the total computing power of a cloud computing task, comprising: Obtain the neural network model corresponding to the cloud computing task; Determine a computational graph corresponding to the neural network model, wherein the computational graph is composed of a plurality of computational graph elements, at least a portion of the plurality of computational graph elements can be operated in parallel, and each of the computational graph elements is any one of the following: a node, an edge between nodes; and determine a total computational amount corresponding to the neural network model during operation based on the parallel computational graph elements included in the computational graph.

15. An electronic device, comprising: A memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement the method of any one of claims 1 to 14.

16. A computer program product comprising: A computer program, which, when executed by a processor of an electronic device, causes the processor to perform the steps of the method of any one of claims 1 to 14.

Images