Node-level execution location decision-making methods, systems, and devices for AI workflows

Abstract

This invention discloses a method, system, and device for node-level execution location decision-making in artificial intelligence workflows. The method parses the AI ​​workflow specifications to obtain currently ready nodes; it obtains an execution profile of each node, which includes at least its data sensitivity requirements and model capability requirements; based on the execution profile, it performs hard constraint filtering to eliminate candidate physical execution positions that do not meet the data sensitivity or model capability requirements, resulting in remaining candidate physical execution positions; it then scores these remaining candidate physical execution positions using a soft constraint scoring formula that includes multi-dimensional evaluation indicators; based on the scoring results, it determines the target execution position of the node and generates scheduling instructions to control the dispatch of the corresponding task to the target execution position for execution. This invention achieves fine-grained computing power routing at the single-node granularity, ensures compliance with underlying data privacy, and improves task success rate and scheduling transparency in complex physical environments.

Description

Technical Field

[0001] This invention belongs to the fields of distributed computing, artificial intelligence workflow orchestration and computing power scheduling, and particularly relates to a method, system and device for node-level execution location decision for artificial intelligence workflow. Background Technology

[0002] With the rapid evolution of Large Language Models (LLM) and AI Agent technologies, orchestrating multiple complex AI tasks into directed acyclic graphs (DAGs) has become the mainstream approach for enterprise applications. To support these complex workflows, the industry has gradually explored distributed execution architectures encompassing the cloud, edge, and terminal devices. However, the inventors have discovered that existing workflow platforms still rely on traditional cloud computing resource scheduling logic in their core "computing power scheduling and execution location decision" mechanism, exhibiting significant technical deficiencies and application bottlenecks, specifically in the following aspects:

[0003] The scheduling granularity is coarse, lacking differentiated routing capabilities based on the characteristics of AI workflow nodes: Existing workflow platforms typically adopt a coarse-grained scheduling model, that is, using the "entire workflow" as the basic scheduling unit and fixing it to a certain execution environment (such as full cloud execution). Existing technologies cannot perform differentiated computing power routing based on the characteristics of "individual nodes" (such as: local data reading nodes vs. global inference and analysis nodes), resulting in a mismatch of computing resources.

[0004] The scheduling strategy is rigid and lacks awareness and coordination regarding data sovereignty and physical context: Existing general-purpose cluster schedulers mainly determine the scheduling location of tasks based on whether resources are idle. This strategy completely ignores the multi-dimensional attribute constraints unique to AI tasks and lacks a systematic awareness of "on-device local data sovereignty," causing data anonymization operations that could be completed locally on the terminal to be forcibly sent to the cloud, leading to serious data leakage and compliance risks. In addition, existing schedulers cannot comprehensively consider dynamic physical context such as terminal battery level and packet loss rate.

[0005] Scheduling decisions are black-boxed, lacking explainability and compliance audit logs: In enterprise-level compliance scenarios, the traditional distributed scheduling process is a "black box." When data flows, the platform cannot provide structured decision-making basis, cannot explain to regulatory authorities "why a node containing private data was routed to the cloud or the edge for execution," and thus fails to meet security audit requirements.

[0006] Lack of fine-grained high availability and degradation mechanisms: Existing workflow scheduling often adopts a static approach of binding a single node. Once the selected execution node fails, the entire workflow fails. There is a lack of node-level alternative link generation and automatic route degradation mechanisms, and the degradation process can easily violate the original security and compliance bottom line, resulting in poor system robustness. Summary of the Invention

[0007] The purpose of this invention is to provide a node-level execution location decision-making method, system, and device for artificial intelligence workflows, in order to solve the technical problems of existing AI workflow scheduling having coarse granularity, weak privacy protection, and lack of dynamic adaptability and interpretability.

[0008] To achieve the above objectives, this invention provides a node-level execution location decision method for artificial intelligence workflows, applied to computing devices performing execution scheduling functions, comprising the following steps:

[0009] Analyze the AI ​​workflow specifications to obtain the nodes that are currently ready to be executed;

[0010] Obtain the execution profile of the node to be executed, wherein the execution profile includes at least the data sensitivity requirements and model capability requirements of the node to be executed;

[0011] Hard constraint filtering is performed based on the execution profile to eliminate candidate physical execution positions that do not meet the data sensitivity requirements or the model capability requirements, resulting in the remaining candidate physical execution positions. The initial candidate physical execution positions include cloud servers, edge gateways, and terminal devices.

[0012] Real-time physical state data of the remaining candidate physical execution positions are collected, and the scheduling priority value of each remaining candidate physical execution position is calculated through a multi-dimensional resource adaptation calculation mechanism. The multi-dimensional resource adaptation calculation mechanism covers at least five technical dimensions: privacy adaptation, latency adaptation, cost adaptation, capability adaptation, and continuity adaptation. The calculation process includes: obtaining the dynamic weight coefficients corresponding to each technical dimension, and weighting them in combination with the quantized feature values ​​of each remaining candidate physical execution position based on the real-time physical state data under different technical dimensions to obtain the scheduling priority value.

[0013] The target execution position of the node to be executed is determined based on the scheduling priority value, and a scheduling instruction is generated to control the dispatch of the corresponding task to the target execution position for execution.

[0014] On the other hand, this invention provides a cloud-edge-device collaborative artificial intelligence workflow node-level scheduling and execution method, applied to a collaborative system including cloud servers, edge gateways, and terminal devices, comprising:

[0015] The scheduling device executes the node-level execution position decision method described above to determine the target execution position of the node to be executed.

[0016] If the target execution point is the terminal device, then the task package containing the business logic of the node to be executed will be sent to the terminal device so that the terminal device can perform privacy data desensitization and local inference calculation in the built-in privacy sandbox, ensuring that the original data does not leave the terminal.

[0017] If the target execution location is the cloud server, the task package or inference request is sent to the edge layer gateway so that the edge layer gateway can determine the tenant identity based on the platform access credentials, and read or reference the upstream model call credentials and model access address corresponding to the tenant identity from the security credential storage module, and proxy route the inference request to the heterogeneous computing power pool or target model service in the cloud server.

[0018] Furthermore, this invention also provides a node-level execution location decision-making system for artificial intelligence workflows, comprising:

[0019] The parsing module is used to parse the AI ​​workflow specifications and obtain the nodes that are currently in a ready state and waiting to be executed.

[0020] The profile acquisition module is used to acquire the execution profile of the node to be executed, and the execution profile includes at least the data sensitivity requirements and model capability requirements of the node to be executed.

[0021] The hard constraint filtering module is used to perform hard constraint filtering based on the execution profile, and remove candidate physical execution positions that do not meet the data sensitivity requirements or the model capability requirements to obtain the remaining candidate physical execution positions. The initial candidate physical execution positions include cloud servers, edge gateways and terminal devices.

[0022] The resource adaptation calculation module is used to collect real-time physical state data of the remaining candidate physical execution positions and calculate the scheduling priority value of each of the remaining candidate physical execution positions through a multi-dimensional resource adaptation calculation mechanism. The multi-dimensional resource adaptation calculation mechanism covers at least five technical dimensions: privacy adaptation, latency adaptation, cost adaptation, capability adaptation, and continuity adaptation. The calculation process includes: obtaining the dynamic weight coefficients corresponding to each technical dimension, and combining them with the quantized feature values ​​of each of the remaining candidate physical execution positions under different technical dimensions based on the real-time physical state data for weighted processing to obtain the scheduling priority value.

[0023] The scheduling output module is used to determine the target execution position of the node to be executed based on the scheduling priority value, and generate scheduling instructions to control the corresponding task to be sent to the target execution position for execution.

[0024] In another aspect, the present invention also provides an electronic device and a computer-readable storage medium.

[0025] The beneficial technical effect of this invention lies in decentralizing the scheduling granularity to a "single node." By improving data privacy compliance, it ensures that highly sensitive nodes are forcibly retained on the edge, linking with the underlying architecture to achieve "data not leaving the edge," reducing the compliance risks caused by the original privacy data leaving the terminal device. Addressing the conflict between high sensitivity and large model capability requirements, under preset scheduling strategies and workflow execution paths, it prevents the original privacy data from being uploaded to the cloud server with inference requests. This invention's technical solution overcomes the limitations of traditional static weights through environment-adaptive dynamic scheduling, introducing environmental context vectors including dimensions such as power consumption and packet loss rate to dynamically calculate resource adaptation weights. Even in weak network and heterogeneous device environments, it maintains a high success rate and lower average latency. A high-availability routing degradation mechanism outputs execution links with degradation semantics, supporting state machine circuit breaking and automatic switching along subsequent alternative links when execution fails. It strictly ensures that degradation links are generated only from legal candidate positions after hard constraint filtering, reducing the risk of breaking privacy or capability constraints during degradation execution and significantly improving the robustness of enterprise-level workflows. Transparent and explainable compliance audit: The pioneering structured reason tags output clearly explains the triggering conditions and weight contributions of execution bit selection, providing enterprises with compliance audit logs that are readable by both machines and humans to cope with data security audits. Attached Figure Description

[0026] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of the invention. Wherein:

[0027] Figure 1 A flowchart of two-stage decision funnel data processing based on execution profile provided in an embodiment of the present invention;

[0028] Figure 2 This is a global scheduling and physical execution interaction diagram of a cloud-edge-device collaborative AI workflow provided in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the user interface for interpretable structured reason tags provided in an embodiment of the present invention. Detailed Implementation

[0030] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0031] The technical solution of the present invention is illustrated below through specific embodiments. The "Node-level Execution Location Decision Method for Artificial Intelligence Workflow" proposed in this invention runs on a computing device with scheduling capabilities (such as a cloud server 100 or an edge gateway 200). This method performs a pre-evaluation of each ready node in the workflow to determine its optimal physical execution location.

[0032] Example 1: A two-stage scheduling decision-making method based on node profiling (core algorithm)

[0033] refer to Figure 1 This demonstrates the internal algorithm data processing flow of the core scheduling brain of this invention. When the workflow reaches a certain frontier node, the scheduling device executes the following decision tree logic: Step S101 (input parsing step): Parse the compiled artificial intelligence workflow specification (WorkflowSpec) and obtain the currently ready nodes to be executed through DAG topology sorting.

[0034] Step S102 (Profile Construction Step): Dynamically construct the execution profile of the node to be executed. Static metadata includes: data_sensitivity (data sensitivity requirement, such as high / medium / low) and model_requirement (model capability requirement, such as large model / small model). Dynamic telemetry data is constructed by extracting the environmental context vector C = [B, N, L] of the underlying physical device in real time, where B is the power status of the terminal device (such as the remaining power percentage), N is the network packet loss rate, and L is the remaining Service Level Agreement (SLA) time margin.

[0035] Step S103 (Hard Constraint Filtering Step): Implement a veto mechanism for the initial cloud server 100, edge gateway 200, and terminal device 300:

[0036] (1) Privacy circuit breaker: If the data_sensitivity in the profile is high and it depends on the local original data, in order to protect data sovereignty, cloud server 100 and edge gateway 200 will be forcibly removed, and only terminal device 300 will be retained.

[0037] (2) Computing power threshold: If model_requirement is a model with very large parameters and the computing power of the terminal and edge sides is lower than the threshold, then 300 terminal devices and 200 edge layer gateways will be removed, and 100 cloud layer servers will be retained.

[0038] (3) Network connectivity: If the device is offline and the node must access the remote model, the cloud side and edge side are not available for the time being.

[0039] Step S104 (Resource Adaptation Dynamic Calculation Step): For the remaining candidate physical execution bits after hard constraint filtering, collect their underlying real-time physical state data, and use a multi-dimensional resource adaptation calculation mechanism to calculate priorities. The specific calculation model is as follows:

[0040] Priority(p)=w1·PrivacyFit(p)+w2·LatencyFit(p)+w3·CostFit(p)+w4·CapabilityFit(p)+w5·ContinuityFit(p)

[0041] Where: p represents the valid candidate physical execution bit retained after hard constraint filtering;

[0042] PrivacyFit(p) represents the privacy fit quantified based on device isolation status. It's important to emphasize that, unlike the veto power in the hard constraint phase, PrivacyFit(p) here is used to handle nodes that "have not triggered hard constraints but still exhibit privacy preferences" (e.g., medium-sensitivity data is preferentially executed via the edge gateway, while low-sensitivity data can be executed via the cloud server to obtain corresponding feature enhancements).

[0043] LatencyFit(p) is the latency fit degree based on network transmission state quantization;

[0044] CostFit(p) is the cost fit degree quantified based on the state of computing resource consumption;

[0045] CapabilityFit(p) is the capability fit quantified based on model computational capabilities;

[0046] ContinuityFit(p) is the continuity fit based on the quantization of the weak network recovery state;

[0047] The five Fit metrics mentioned above are all normalized to values ​​in the range [0,1], serving as quantitative feature values ​​for each technical dimension.

[0048] w1 to w5 are dynamic weight coefficients, and they satisfy w1+w2+w3+w4+w5=1. The system dynamically adjusts the weight coefficients based on the environmental context vector C = (B, N, L). For example: when the remaining power B of the terminal device is lower than the preset threshold, the quantization characteristic value of the terminal device on CostFit(p) is reduced, or the penalty weight of energy consumption cost in CostFit(p) is increased (to exclude local high power consumption inference); when the network packet loss rate N increases, the weight of weak network continuity w5 is increased; when the SLA time margin L decreases, the latency weight w2 is increased.

[0049] Step S105 (Link Generation Step): Arrange the execution links in descending order of scheduling priority values ​​to generate execution links (e.g., [Cloud Layer 100, Edge Layer 200, End Layer 300]) containing the main execution position (target execution position) and ordered alternative execution positions. Extract the decision trajectory to generate structured cause labels and finally generate scheduling instructions for issuance.

[0050] Example 2: Cloud-edge-device collaborative physical execution closed loop based on scheduling decisions;

[0051] refer to Figure 2 This embodiment details how the "target execution bit" output by the algorithm in Embodiment 1 controls the task flow of the underlying physical device. Scenario A: Local data digest node (edge-side execution mechanism)

[0052] A workflow node needs to read a user's local document and extract a summary. In step S103 of Example 1, the data is identified as having extremely high sensitivity, triggering a privacy circuit breaker, and the decision target execution bit is terminal device 300.

[0053] The scheduling device distributes the task package to the terminal device 300. The terminal device 300 then launches its local execution engine 310 within its internal privacy sandbox 320, performing sensitive data reading, de-identification, and lightweight model inference locally. This scheduling decision enhances the robustness and auditability of workflow node execution, achieving "data not leaving the terminal."

[0054] Scenario B: Global Comprehensive Analysis Node (Cloud-based Execution Mechanism)

[0055] A certain node needs to summarize the results of multiple chapters and call a high-capability model to output the final solution. Step S103 determines that it depends on the anonymized intermediate results (low sensitivity) and has high requirements for model capabilities. The decision target execution position is cloud server 100.

[0056] The task package is distributed to the edge layer gateway 200, which determines the tenant identity based on the platform access credentials. After reading or referencing the upstream model call credentials and model access address corresponding to the tenant identity from the security credential storage module, the edge layer gateway 200 dynamically routes the request to the heterogeneous computing power pool or target model service in the cloud server 100 for high-performance inference. The edge layer gateway 200 is only responsible for result aggregation and callback forwarding.

[0057] Example 3: High-availability state machine degradation mechanism based on alternative links (weak network office scenario);

[0058] Enterprise-level AI workflows have extremely high robustness requirements. The scheduling algorithm of this invention not only outputs a single node, but also a complete "execution chain," thereby providing physical-level routing degradation protection.

[0059] State machine circuit breaker and switching logic: If the system detects a network interruption or computing power overflow exception in the main execution unit (such as cloud server 100), the system will not throw a global failure, but will trigger the state machine circuit breaker. The system will intercept the exception and automatically redirect the input context to the next ordered alternative execution unit in the execution chain generated in step S105 (such as downgrading to edge gateway 200 or terminal device 300) to re-initiate execution, ensuring business continuity.

[0060] Specific application scenarios (weak network environment):

[0061] When a user uses terminal device 300 in a weak network environment such as high-speed rail, the scheduling equipment extracts the context vector and detects that the current network packet loss rate is extremely high.

[0062] During the soft constraint scoring (S104) phase, the weight of weak network recoverability (ContinuityFit) is dynamically amplified. The system prioritizes allocating the main execution bit of "simple nodes that can be executed offline" to the terminal device 300 for cached execution; "complex nodes that must be connected to the network" are suspended through alternative links or proxied by the nearest edge layer gateway 200. After the network is restored, subsequent frontier nodes will continue to be advanced.

[0063] Example 4: Compliance audit logs and structured reason tags;

[0064] refer to Figure 3 This embodiment demonstrates the visual interaction process of the structured cause labels output by the present invention on the console, solving the problem of the black box nature of traditional scheduling and enabling the platform to have the auditing capability to explain "why a certain node is executed on the cloud side or the edge side". Figure 3As shown, the user interface is presented in a left-right split-column structure. The left side of the interface is the workflow canvas area 401, which displays the directed acyclic graph of the AI ​​workflow, in which a node that has completed scheduling is highlighted.

[0065] The right side of the interface displays a pop-up node scheduling details panel 402. This panel directly parses and renders the data output from algorithm step S105. The structured reason label display area 403 clearly shows:

[0066] (1) Main execution bit: Terminal device 300;

[0067] (2) Alternative links: None;

[0068] (3) Decision Reason Label: The hard constraint identifiers that were triggered are displayed in a structured list (e.g., [!] identifier: 0x01) with an explanatory text: "Data sensitivity = high, contains personal privacy, the system has forced it to be isolated and executed in the local sandbox, and the cloud route has been circuit-broken."

[0069] pass Figure 3 The interface of this invention not only outputs the execution location, but also the structured decision reasons, providing intuitive and reliable data flow evidence for A / B optimization, failure tracking and security compliance auditing of enterprise-level AI platforms.

[0070] Hardware and Storage Medium Embodiments: According to embodiments of the present invention, the present invention also provides an electronic device, a readable storage medium, and a computer program product.

[0071] The example electronic devices of the embodiments of the present invention are intended to represent various forms of digital computers or network computing devices, including computing units (processors) that can perform various appropriate actions and processes according to computer programs stored in read-only memory (ROM) or computer programs loaded from storage units into random access memory (RAM) to implement the node-level execution location decision method described in embodiments 1 to 4 above.

[0072] RAM can also store various programs and data required for the operation of electronic devices. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus. Multiple components in the electronic device are connected to the I / O interface, including: input units (such as keyboards and mice), output units (such as displays), storage units (such as disks), and communication units (such as network cards). The communication unit allows the electronic device to exchange information / data with other devices through computer networks. The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A node-level execution location decision method for artificial intelligence workflows, applied to computing devices performing execution scheduling functions, characterized in that, Includes the following steps: Analyze the AI ​​workflow specifications to obtain the nodes that are currently ready to be executed; Obtain the execution profile of the node to be executed, wherein the execution profile includes at least the data sensitivity requirements and model capability requirements of the node to be executed; Hard constraint filtering is performed based on the execution profile to eliminate candidate physical execution positions that do not meet the data sensitivity requirements or the model capability requirements, resulting in the remaining candidate physical execution positions. The initial candidate physical execution positions include cloud servers, edge gateways, and terminal devices. Real-time physical state data of the remaining candidate physical execution positions are collected, and the scheduling priority value of each remaining candidate physical execution position is calculated through a multi-dimensional resource adaptation calculation mechanism. The multi-dimensional resource adaptation calculation mechanism covers at least five technical dimensions: privacy adaptation, latency adaptation, cost adaptation, capability adaptation, and continuity adaptation. The calculation process includes: obtaining the dynamic weight coefficients corresponding to each technical dimension, and weighting them in combination with the quantized feature values ​​of each remaining candidate physical execution position based on the real-time physical state data under different technical dimensions to obtain the scheduling priority value. The target execution position of the node to be executed is determined based on the scheduling priority value, and a scheduling instruction is generated to control the dispatch of the corresponding task to the target execution position for execution.

2. The method according to claim 1, characterized in that, The execution of hard constraint filtering based on the execution profile includes: If the data sensitivity requirement in the execution profile is higher than the preset level and the node to be executed depends on local raw data, then the cloud server and edge gateway will be marked as not meeting the hard privacy constraints. If the model capability requirement in the execution profile is higher than the computing power threshold of the terminal device and the edge layer gateway, the terminal device and the edge layer gateway are marked as not meeting the capability hard constraint; when the same node to be executed triggers both the privacy hard constraint and the capability hard constraint, resulting in no single candidate physical execution position that meets all hard constraints, the node splitting process is triggered.

3. The method according to claim 1, characterized in that, The process of obtaining the dynamic weight coefficients corresponding to each technical dimension includes: Real-time extraction of the environmental context vector of the underlying physical device, the environmental context vector including the terminal device's power status, network quality status, and service level agreement time margin; The dynamic weight coefficients corresponding to each technical dimension are dynamically adjusted based on the environmental context vector, or the quantization feature values ​​of the corresponding candidate physical execution bits under a specific technical dimension are dynamically adjusted based on the environmental context vector.

4. The method according to claim 3, characterized in that, The method of dynamically adjusting the dynamic weight coefficients corresponding to each technical dimension based on the environmental context vector, or dynamically adjusting the quantization feature value of the corresponding candidate physical execution bit under a specific technical dimension based on the environmental context vector, includes: when the battery status of the terminal device is lower than a preset threshold, reducing the quantization feature value of the terminal device in cost adaptability, or increasing the penalty weight of energy consumption cost in cost adaptability; mapping the packet loss rate in the network quality status to the weight coefficient of the continuity adaptability in a positive correlation; and normalizing all weight coefficients calculated by the dynamic mapping as the dynamic weight coefficients.

5. The method according to claim 1, characterized in that, Determining the target execution position of the node to be executed based on the scheduling priority value includes: Based on the descending order of the scheduling priority values, an execution chain is generated, which includes a main execution position and ordered candidate execution positions, and the main execution position is used as the target execution position; the ordered candidate execution positions are generated only based on the remaining candidate physical execution positions retained after the hard constraint filtering. The method further includes a degradation step based on the execution link: After the task is issued, monitor the execution status of the main execution bit; If a network interruption or computing power anomaly is detected, the state machine circuit breaker is triggered, and the input context is redirected to the next ordered candidate execution position in the execution link for physical degradation execution, so that the node does not break the privacy constraints, model capability constraints and link connectivity constraints determined by the hard constraint filtering step during degradation execution.

6. The method according to claim 1, characterized in that, The method further includes: During the execution of the hard constraint filtering and the resource adaptation calculation, the condition identifier that triggers the filtering rule and the maximum weight contribution item in the calculation process are extracted. The condition identifier and the maximum weight contribution item are assembled into a structured cause label; The structured reason label is output synchronously with the scheduling instruction to provide a compliance audit log explaining the reasons for the target execution bit selection.

7. A cloud-edge-device collaborative AI workflow node-level scheduling and execution method, applied to a collaborative system including cloud servers, edge gateways, and terminal devices, characterized in that... include: The scheduling device executes the node-level execution position decision method as described in any one of claims 1 to 6 to determine the target execution position of the currently to-be-executed node; If the target execution point is the terminal device, then the task package containing the business logic of the node to be executed will be sent to the terminal device so that the terminal device can perform privacy data desensitization and local inference calculation in the built-in privacy sandbox, ensuring that the original data does not leave the terminal. If the target execution location is the cloud server, the task package or inference request is sent to the edge layer gateway so that the edge layer gateway can determine the tenant identity based on the platform access credentials, and read or reference the upstream model call credentials and model access address corresponding to the tenant identity from the security credential storage module, and proxy route the inference request to the heterogeneous computing power pool or target model service in the cloud server.

8. The method according to claim 7, characterized in that, The node splitting process includes: the method also includes a step to handle conflicts between privacy constraints and model capability constraints. When the node to be executed simultaneously meets the conditions that the data sensitivity is higher than the preset level and the model capability requirement is higher than the terminal device computing power threshold, the node to be executed is split into a local privacy preprocessing sub-node and a cloud model inference sub-node. The local privacy preprocessing sub-node is scheduled to be executed on the terminal device to perform privacy identification and de-identification processing on the local raw data; The cloud-based model inference sub-node is scheduled to be executed on the cloud server for model inference based on the anonymized intermediate results.

9. The method according to claim 7, characterized in that, If the target execution node is the terminal device, then the task package containing the business logic of the node to be executed is sent to the terminal device, including: The system identifies whether a user-provided external AI agent tool is registered on the terminal device. If so, the task package is converted into a dual semantic format instruction containing structured data and markup language text, and then sent to the external AI agent tool for local execution via a command-line interface.

10. A node-level execution location decision-making system for artificial intelligence workflows, characterized in that, include: The parsing module is used to parse the AI ​​workflow specifications and obtain the nodes that are currently in a ready state and waiting to be executed. The profile acquisition module is used to acquire the execution profile of the node to be executed, and the execution profile includes at least the data sensitivity requirements and model capability requirements of the node to be executed. The hard constraint filtering module is used to perform hard constraint filtering based on the execution profile, and remove candidate physical execution positions that do not meet the data sensitivity requirements or the model capability requirements to obtain the remaining candidate physical execution positions. The initial candidate physical execution positions include cloud servers, edge gateways and terminal devices. The resource adaptation calculation module is used to collect real-time physical state data of the remaining candidate physical execution positions and calculate the scheduling priority value of each of the remaining candidate physical execution positions through a multi-dimensional resource adaptation calculation mechanism. The multi-dimensional resource adaptation calculation mechanism covers at least five technical dimensions: privacy adaptation, latency adaptation, cost adaptation, capability adaptation, and continuity adaptation. The calculation process includes: obtaining the dynamic weight coefficients corresponding to each technical dimension, and combining them with the quantized feature values ​​of each of the remaining candidate physical execution positions under different technical dimensions based on the real-time physical state data for weighted processing to obtain the scheduling priority value. The scheduling output module is used to determine the target execution position of the node to be executed based on the scheduling priority value, and generate scheduling instructions to control the corresponding task to be sent to the target execution position for execution.

11. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program to implement the method according to any one of claims 1 to 8.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 8.

Images