Information processing method and device, electronic equipment and storage medium

Abstract

The present disclosure provides an information processing method and device, electronic equipment and storage medium, and relates to the technical field of artificial intelligence and computer. The specific implementation scheme is: during inference task execution, log information associated with each of a plurality of life cycle nodes is obtained by calling a target interface; wherein the plurality of life cycle nodes indicate the logical relationship between a plurality of operations in the inference task execution process; and each log information is stored in association according to the logical relationship, so as to trace the execution path of the inference task based on each log information.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and computer technology, and in particular to information processing methods, apparatus, electronic devices, and storage media. Background Technology

[0002] Conducting compliance audits on inference tasks for privately deployed AI platforms can improve the credibility of inference task execution results and reduce uncontrollable risks.

[0003] However, the lack of structured time recording capabilities for the complete execution chain of inference tasks makes it difficult to meet the need for tracing the execution path of inference tasks. Summary of the Invention

[0004] This disclosure provides an information processing method, apparatus, electronic device, and storage medium.

[0005] According to one aspect of this disclosure, an information processing method is provided, comprising: during the execution of an inference task, obtaining log information associated with multiple lifecycle nodes by calling a target interface; wherein the multiple lifecycle nodes indicate the logical relationship between multiple operations during the execution of the inference task; and storing the log information in association according to the logical relationship, so as to trace the execution path of the inference task based on the log information.

[0006] According to another aspect of this disclosure, an information processing apparatus is provided, including: an acquisition module and a storage module.

[0007] The acquisition module is used to obtain log information associated with multiple lifecycle nodes by calling the target interface during the execution of the inference task; the multiple lifecycle nodes indicate the logical relationship between multiple operations in the execution process of the inference task.

[0008] The storage module is used to associate and store log information according to logical relationships, so as to trace the execution path of inference tasks based on each log information.

[0009] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0010] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform the methods described above.

[0011] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method described above.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0014] Figure 1 This illustration schematically shows an exemplary system architecture to which information processing methods and apparatus can be applied according to embodiments of the present disclosure;

[0015] Figure 2 A flowchart illustrating an information processing method according to an embodiment of the present disclosure is shown schematically.

[0016] Figure 3 The diagram illustrates the interaction between modules for performing an information processing method according to an embodiment of the present disclosure.

[0017] Figure 4 This illustration shows a diagram of how log information of multiple lifecycle nodes are stored in association according to logical relationships, based on an embodiment of the present disclosure.

[0018] Figure 5 The illustration shows a schematic diagram of log querying and trace analysis for inference tasks according to embodiments of the present disclosure;

[0019] Figure 6 A block diagram of an information processing apparatus according to embodiments of the present disclosure is schematically shown; and

[0020] Figure 7 A block diagram of an electronic device suitable for implementing an information processing method according to an embodiment of the present disclosure is shown schematically. Detailed Implementation

[0021] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0022] When addressing the auditing requirements for inference tasks on privately deployed AI platforms, the relevant examples mainly suffer from the following drawbacks:

[0023] (1) Logs are recorded at the system call level, which is disconnected from the business semantics of the inference task. At least tens of thousands of system calls are involved in a single execution of an inference task, which increases the difficulty of extracting information that meets audit requirements.

[0024] (2) The large language model watermarking technology only covers the verification of the ownership of the output content, making it difficult to trace the status of other nodes involved in the execution of the reasoning task.

[0025] (3) The tracing method based on the model training phase is difficult to cover the log data generated during the execution of the inference task. Furthermore, since this method relies on the blockchain consensus mechanism, it lacks a reliable deployment path in the private enterprise intranet environment, and the consensus latency is difficult to meet the real-time auditing requirements of high-throughput inference scenarios.

[0026] (4) By injecting tracing identifiers during service calls, the latency and path of service calls are recorded. There is a lack of semantic modeling of the lifecycle of AI inference tasks, and the fields are geared towards service performance analysis rather than auditing requirements. Therefore, it is difficult to guarantee data integrity, log records can be arbitrarily modified, and it is difficult to meet the requirement of immutability of audit evidence.

[0027] In view of this, the embodiments of this disclosure, during the execution of the inference task, obtain the log information of each of the multiple lifecycle nodes in the inference task execution process by calling the target interface, which significantly improves the traceability of the inference task execution path. Simultaneously, according to the logical relationships between the multiple operations in the inference task execution process indicated by the multiple lifecycle nodes, the multiple log information is stored in association, transforming the logs generated by each lifecycle node during the inference task execution into an irreversible causal link, further improving the completeness of the log information and the level of information security.

[0028] Figure 1 The illustration schematically depicts an exemplary system architecture to which information processing methods and apparatus can be applied according to embodiments of the present disclosure.

[0029] It is important to note that Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments, or scenarios. For example, in another embodiment, an exemplary system architecture to which information processing methods and apparatus can be applied may include a terminal device, but the terminal device may implement the information processing methods and apparatus provided by the embodiments of this disclosure without interacting with a server.

[0030] like Figure 1As shown, the system architecture 100 according to this embodiment may include: an AI (Artificial Intelligence) inference platform 110, a downstream system 120, a storage area 130, and an LLM (Large Language Model) 140.

[0031] When the AI ​​inference platform 110 receives an inference task, it can execute the inference task according to multiple lifecycle nodes of the inference task. For example, node E1 can represent the task submission node, node E2 can represent the task queuing node, node E3 can represent the resource allocation node, node E4 can represent the model loading node, node E5 can represent the inference execution node, node E6 can represent the output delivery node, and node E7 can represent the determination node. The logs generated by each node are appended and associated with each other and stored in the storage area 130.

[0032] Downstream system 120 represents the application end that receives the result data output by the AI ​​inference platform 110 during the execution of the inference task. This confirmation node is triggered when the AI ​​inference platform 110 completes the inference task and transmits the result data to the downstream system 120, and the downstream system confirms receipt of the result data and applies the result data.

[0033] Because logs for each lifecycle node during the execution of the inference task are stored in storage area 130, when auditing is required, the log information sequence of the inference task can be queried from the storage area based on audit requirements. After verifying the information integrity, the logs are input into LLM140 to perform a traceability analysis of the inference task execution path, obtaining information that records the complete link from the submission of the inference task to the downstream system application.

[0034] In the technical solutions disclosed herein, the collection, storage, use, processing, transmission, provision, disclosure, and application of any type of information, such as user personal information, comply with the provisions of relevant laws and regulations, necessary confidentiality measures have been taken, and they do not violate public order and good morals.

[0035] In the technical solution disclosed herein, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0036] Figure 2 A flowchart illustrating an information processing method according to an embodiment of the present disclosure is shown schematically.

[0037] like Figure 2 As shown, the method 200 may include operations S210~S220.

[0038] During operation S210, in the execution of the inference task, log information associated with multiple lifecycle nodes is obtained by calling the target interface.

[0039] In operation S220, log information is associated and stored according to logical relationships, so as to trace the execution path of the reasoning task based on the log information.

[0040] The target interface can be an interface that can call lifecycle hook functions, which automatically trigger log information collection operations at each lifecycle node during the execution of the inference task to obtain the log information associated with each lifecycle node.

[0041] For example, an interceptor chain pattern can be used. Without modifying the code of the AI ​​intelligent inference platform used to execute inference tasks, interceptors corresponding to each lifecycle node are registered with the inference task scheduler in a non-intrusive manner, specifying the triggering time for each interceptor. For instance, the interceptor corresponding to the task submission node can be triggered when the task scheduler receives the inference task submission request. During the inference task state transition, the post-hook method of the corresponding lifecycle node's interceptor can be called to obtain the log information of the inference task at the current lifecycle node. Since the interceptors can directly read the context information including the execution phase of the inference task, there is no need to modify the function signature or calling convention of the task scheduler, thus achieving non-intrusive integration.

[0042] In addition, the interceptor can store the log information of each lifecycle node as an audit object in a memory queue, reducing the additional latency to the main link of the task scheduler. For example, the additional latency of the main link can be less than 0.5ms, reducing the impact of intercepting the log information of each lifecycle node on the normal execution of the inference task.

[0043] In some embodiments, in a scenario processing 100 inference tasks per second, the average additional latency introduced by the interceptor can be less than 0.3ms. When the memory queue is close to full load, a backpressure strategy can be triggered, degrading to a mode that collects log information at a fixed ratio and triggering a full load warning.

[0044] In some embodiments, multiple lifecycle nodes may include task state transition nodes along the complete execution path of the inference task, from task submission to the application of the result data. Examples include: task submission node, task queuing node, resource allocation node, model loading node, inference execution node, output delivery node, and determination node.

[0045] Therefore, multiple lifecycle nodes indicate the logical relationships between multiple operations during the execution of the inference task.

[0046] Because each log entry in a traditional logging system is stored independently, any log entry can be maliciously modified without affecting other log entries, thus causing a break in the information chain.

[0047] Therefore, the embodiments of this disclosure can associate and store log information according to logical relationships, and retain the causal relationship between log information during the storage process, thereby forming an irreversible causal information chain, providing an effective verification mechanism for tracing the execution path of reasoning tasks.

[0048] For example, log messages can be appended only and logically linked into a cryptographically tamper-proof hash chain structure, which is then written into the storage area. For instance, the log message generated by the nth lifecycle node can be concatenated with the log messages generated by the previous n-1 lifecycle nodes, and then hashed to obtain the hash chain structure, where n is an integer greater than 1.

[0049] This storage area does not need to provide any interface for modifying data. Any modification operation on the log information in the storage area is completed by appending new log information.

[0050] In AI platform scenarios for private deployment, the hash chain structure can be partitioned and stored based on user identifiers according to actual needs. The hash chain structures of different users are independent of each other to support the audit isolation requirements of multi-user data.

[0051] In some embodiments, when auditing of inference tasks is required, log information can be read from the storage area and the log information can be verified based on the hash chain structure to determine whether it has been tampered with. If the verification is successful, the execution path of the inference task can be traced based on the log information using a large language model.

[0052] According to embodiments of this disclosure, during the execution of an inference task, log information of multiple lifecycle nodes during the inference task execution process is obtained by calling a target interface, significantly improving the traceability of the inference task execution path. Simultaneously, based on the logical relationships between multiple operations during the inference task execution process indicated by the multiple lifecycle nodes, the multiple log information is stored in association, transforming the logs generated by each lifecycle node during the inference task execution into an irreversible causal chain, further improving the completeness and security of the log information.

[0053] According to embodiments of this disclosure, multiple lifecycle nodes include: a task submission node; and log information including first log information. During the execution of an inference task, obtaining the log information associated with each of the multiple lifecycle nodes by calling a target interface may include the following operation: in response to receiving a submission request for an inference task, obtaining the first log information associated with the task submission node by calling the target interface.

[0054] In this embodiment of the disclosure, the first log information indicates information about the inference task. Exemplarily, the inference task information includes, but is not limited to, a task identifier field, a user identifier field that submitted the task, and a submission time field. In some embodiments, a user role field can also be obtained by calling the identity and access management interface provided by the AI ​​intelligent platform. The user role field records the user's role at the time of task submission, such as "approver," which can be used to verify afterward whether the user has authorization to call a specific model or process specific data, supporting internal control and compliance review scenarios.

[0055] The log information associated with the task submission node accurately records the information of the inference task, providing the necessary elements of the responsible party for the traceability of the inference task, and meeting the internal control and compliance audit requirements for privately deployed artificial intelligence platforms.

[0056] In private deployment scenarios, the deployed computing resources are limited, and task queuing may occur when multiple inference tasks are submitted. Therefore, in some embodiments, multiple lifecycle nodes include task queuing nodes. Log information includes fourth log information associated with the task queuing nodes.

[0057] According to embodiments of this disclosure, during the execution of an inference task, obtaining log information associated with multiple lifecycle nodes by calling a target interface may include the following operation: in response to determining that the inference task is in a waiting state, obtaining fourth log information by calling the target interface.

[0058] In this embodiment of the disclosure, the fourth log information indicates the waiting status information of the inference task. Exemplarily, the fourth log information may include, but is not limited to, a field indicating the position of the inference task in the task queue and a field indicating the estimated waiting time.

[0059] For example, if there are 10 inference tasks waiting to be processed in the task queue, and this inference task is the 3rd one, then "3 / 10" can be filled into the position field in the fourth log information.

[0060] The estimated waiting time can be determined based on the current computing resource status of the AI ​​platform and the data processing volume and task complexity of the first two inference tasks in the task queue.

[0061] For example, the specific method for determining the estimated waiting time can include, but is not limited to, pre-configuring the mapping relationship between data processing volume, task complexity, and estimated processing time. By querying the estimated processing time of the two preceding inference tasks, the sum of the estimated processing times of the two preceding inference tasks is determined as the estimated waiting time of the current inference task. This method can minimize the consumption of computing resources during the acquisition of fourth log information and reduce the impact on the execution process of the inference task.

[0062] The log information associated with the task queuing nodes accurately records the waiting status information of the inference tasks, providing evidence for tracing the reasons for the differences in the execution time of the inference tasks and meeting the requirement for complete tracing of the execution path of the inference tasks.

[0063] The execution efficiency of inference tasks is related to the resource allocation status. Therefore, in some embodiments, the log information acquisition operation of the resource allocation node can be triggered when the resource scheduler completes the allocation of GPU (Graphics Processing Unit) nodes.

[0064] According to embodiments of this disclosure, during the execution of an inference task, obtaining log information associated with multiple lifecycle nodes by calling a target interface may include the following operation: in response to determining that the inference task has been assigned to a target processing node, obtaining fifth log information by calling the target interface.

[0065] The target processing node can be any GPU deployed on an artificial intelligence platform for processing inference tasks.

[0066] The fifth log information indicates the resource allocation status information of the inference task. For example, the fifth log information may include, but is not limited to, the identification field of the target processing node and the resource configuration field of the target processing node, the resource configuration field including, but not limited to, the maximum amount of video memory that the target processing node can allocate to an inference task.

[0067] For example, if the resource scheduler assigns an inference task to GPU0, and GPU0 can allocate a maximum of 20GB of video memory to the inference task, then "GPU0" can be filled in the identifier field of the target processing node in the fifth log information, and "20GB" can be filled in the resource configuration field of the target processing node.

[0068] The log information associated with the resource allocation node records the resource allocation status information of the inference task, providing evidence for tracing the reasons affecting the efficiency of task execution, and further improving the completeness of tracing the execution path of the inference task.

[0069] A privately deployed AI platform can be configured with multiple models, including but not limited to large language models, computer vision models, multimodal models, and traditional machine learning models. Different inference tasks may require calling different types of models, and the loading time of different types of models will also affect the overall execution efficiency of the inference task.

[0070] Therefore, in this embodiment of the disclosure, multiple lifecycle nodes may include: a model loading node. Log information includes sixth log information associated with the model loading node.

[0071] According to embodiments of this disclosure, during the execution of an inference task, obtaining log information associated with multiple lifecycle nodes by calling a target interface may include the following operation: in response to determining that the model used to perform the inference task is in a loaded state, obtaining the sixth log information by calling the target interface.

[0072] For example, when it is determined that the model used to perform the inference task is in a loaded state, the operation of obtaining the sixth log information generated for the model loading node is triggered.

[0073] In this embodiment of the disclosure, the sixth log information indicates the model's loading information. The sixth log information may include, but is not limited to, a model identifier field and a loading duration field.

[0074] For example, the inference task could be a multimodal recognition task targeting text and images, and the model matching this inference task could be a multimodal model. Therefore, the model identifier field in the sixth log information could be filled with an identifier field used to indicate the multimodal model.

[0075] The loading time field is used to fill in the time required for the multimodal model to load from the start to the end. For example, if the loading time of the multimodal model is 5 seconds, you can fill in "5 seconds" in the loading time field of the sixth log message.

[0076] The log information associated with the model loading node accurately records the loading information of the model used to perform the inference task, providing auxiliary information for tracing the reasons affecting the efficiency of task execution and further improving the evidence chain of the inference task execution path.

[0077] In private deployment scenarios, especially for industries such as financial risk control, government approval, and energy dispatch, the data that the inference task needs to process are all highly sensitive assets. In order to meet the data security requirements, it is explicitly required that the data must not be transmitted to external systems or third-party auditing platforms.

[0078] Therefore, the log information generated by nodes related to the data to be processed needs to not only meet the requirements of data integrity, but also the requirements of data security.

[0079] In this embodiment of the disclosure, the node associated with the data to be processed may include an inference execution node. Log information may include second log information associated with the inference execution node.

[0080] According to embodiments of this disclosure, during the execution of an inference task, obtaining log information associated with multiple lifecycle nodes by calling a target interface may include the following operations: obtaining first initial log information associated with the inference execution node by calling the target interface; extracting features from the initial data to be processed by the inference task in the memory of the execution process of the inference task to obtain initial data feature information; and writing the initial data feature information into the first initial log information to obtain second log information.

[0081] The first initial log information may include at least an inference task field and an initial data feature field. In the first initial log information, the inference task field may be filled with relevant information about the inference task, and the initial data feature field may be "empty".

[0082] Privately deployed AI platforms all have an independent and secure runtime environment that hosts the inference model used to perform inference tasks. The edge of this runtime environment is defined by the private deployment architecture at the operating system or infrastructure level so that the inference process runs in the memory space of this runtime environment.

[0083] Therefore, in order to meet data security requirements, feature extraction is performed on the initial data to be processed by the inference task in the memory of the execution process of the inference task to obtain the initial data feature information.

[0084] For example, features can be extracted by performing a hash operation on the initial data, and the calculated hash value can be used as the initial data feature information. This initial data feature information can serve as a data fingerprint that uniquely identifies the initial data. For instance, a 256-bit secure hash algorithm can be used to process the initial data to obtain the initial data feature information.

[0085] Then, the initial data feature information is filled into the initial data feature field of the first initial log information to obtain the second log information.

[0086] In the memory of the inference task's execution process, features are extracted from the initial data to be processed, yielding initial data feature information. This initial data feature information can be used for subsequent auditing and tracing, providing cryptographically significant integrity evidence for tracing the execution path of the inference task without transmitting the initial data, thus further enhancing data security.

[0087] In this embodiment of the disclosure, the node associated with the data to be processed may further include an output delivery node. The log information may also include third log information associated with the output delivery node.

[0088] According to embodiments of this disclosure, during the execution of an inference task, obtaining log information associated with multiple lifecycle nodes by calling a target interface may include the following operations: obtaining second initial log information associated with the output delivery node by calling the target interface; extracting features from the result data of the inference task in the memory of the execution process of the inference task to obtain result data feature information; and writing the result data feature information into the second initial log information to obtain third log information.

[0089] The second initial log information may include at least an inference task field and a result data feature field. In the second initial log information, the inference task field may be filled with relevant information about the inference task, while the result data feature field may be "empty".

[0090] To meet data security requirements, feature extraction is performed on the result data of the inference task within the memory of the inference task execution process to obtain the result data feature information.

[0091] The result data can represent the reasoning results obtained by performing a reasoning task on the initial data. For example, the initial data can be a question text described in natural language, and the result data can be the response text generated by a large language model using its natural semantic understanding capabilities.

[0092] For example, feature extraction can be performed by hashing the result data, and the calculated hash value can be used as the result data feature information. This result data feature information can serve as a unique data fingerprint identifying the result data. For instance, a 256-bit secure hash algorithm can be used to process the result data to obtain the result data feature information.

[0093] Then, the result data feature information is filled into the result data feature field of the second initial log information to obtain the third log information.

[0094] In the memory of the inference task's execution process, feature extraction is performed on the result data to obtain result data feature information. This result data feature information can be used for subsequent auditing and tracing, providing cryptographically significant integrity evidence for tracing the execution path of the inference task without transmitting the initial data, thus further improving data security.

[0095] To achieve closed-loop collection of end-to-end information during the execution of the inference task, the multiple lifecycle nodes in this embodiment may further include a determining node. This determining node can be configured in any downstream system used to apply or consume the result data of the inference task. Log information includes seventh log information associated with the determining node.

[0096] According to embodiments of this disclosure, during the execution of an inference task, obtaining log information associated with multiple lifecycle nodes by calling a target interface may include the following operation: in response to receiving a determination operation for the result data of the inference task, obtaining seventh log information by calling the target interface.

[0097] For example, the determination operation may include, but is not limited to, the confirmation operation and / or usage operation of the downstream system for the result data. When confirming or using the result data of the inference task, the downstream system may proactively send the seventh log information generated by the determination node to the target interface.

[0098] The seventh log information indicates the application status information of the result data. This application status information includes, but is not limited to, the identifier field of the downstream system and the field indicating the execution time of the operation.

[0099] The log information associated with the identified node records the application status information of the downstream system to the result data of the inference task. By acquiring the log information generated by the identified node, a closed-loop collection of the entire link information in the execution process of the inference task is realized, which further improves the completeness of the evidence information required to trace the execution path of the inference task.

[0100] It should be noted that the lifecycle nodes described in the above embodiments are exemplary, and the information processing method of this disclosure can be applied to the information processing of lifecycle nodes of reasoning tasks under any reasoning framework, and is not limited to the nodes described in the above embodiments.

[0101] In some embodiments, preprocessing operations may be performed before storing the log information associated with each lifecycle node in order to standardize the format of the log information for each lifecycle node.

[0102] For example, it can verify whether the format of each field in the log information is correct and whether there are any missing fields. Log information with missing fields can be marked as abnormal logs and stored in an abnormal log queue, and will not be included in the tracing operations for inference tasks.

[0103] For example, you can verify whether the format of the model identifier field in the log information is correct in order to eliminate the impact of differences in the same model identifier in different deployment environments on the traceability operation.

[0104] For example, the timestamps of each lifecycle node can be corrected by using the network time protocol calibration value to ensure the time sequence consistency of the entire link log information of the inference task.

[0105] Figure 3 An interactive diagram illustrating the various modules for performing an information processing method according to an embodiment of the present disclosure is shown.

[0106] like Figure 3 As shown, the apparatus for executing the information processing method of the embodiments of this disclosure may include a task scheduler, a probe injection module, a standardization module, an association module, and a storage area. The log information associated with each lifecycle node may be represented as an Audit Event (hereinafter referred to as AE).

[0107] When the task scheduler triggers a task submission event, the probe injection module collects the original AE1 containing information about the inference task by calling the target interface. Then, the normalization module preprocesses the original AE1 to obtain a normalized AE1. Finally, the association module performs a hash operation on the normalized AE1 to obtain the log feature information SHA256(E1) associated with the task submission node E1. SHA256(E1) represents the log feature information associated with the task submission node E1, and SHA256 (Secure Hash Algorithm 256-bit) is used for this purpose. The log feature information SHA256(E1) and the log information associated with the task submission node E1 are then stored in the storage area.

[0108] When the task scheduler triggers a task queuing event, the probe injection module collects and records the raw AE2 of the waiting status information of the inference task by calling the target interface. Then, the normalization module preprocesses the raw AE2 to obtain the normalized AE2, and the association module performs a hash operation on the normalized AE2 to obtain the log feature information SHA256(E2) of the node. SHA256(E2) represents the log feature information associated with the task queuing node E2, and the log feature information SHA256(E2) and the log information associated with the task queuing node E2 are stored in the storage area.

[0109] When the task scheduler triggers a resource allocation event, the probe injection module collects and records the raw AE3 of the resource allocation status information of the inference task by calling the target interface. Then, the normalization module preprocesses the raw AE3 to obtain the normalized AE3, and the association module performs a hash operation on the normalized AE3 to obtain the log feature information SHA256(E3) of the node. SHA256(E3) represents the log feature information associated with the resource allocation node E3, and the log feature information SHA256(E3) and the log information associated with the resource allocation node E3 are stored in the storage area.

[0110] When the task scheduler triggers a model loading event, the probe injection module collects and records the raw AE4 of the model loading information by calling the target interface. Then, the normalization module preprocesses the raw AE4 to obtain a normalized AE4. Finally, the association module performs a hash operation on the normalized AE4 to obtain the log feature information SHA256(E4) associated with that node. SHA256(E4) represents the log feature information associated with the model loading node E4, and both the log feature information SHA256(E4) and the log information associated with the model loading node E4 are stored in the storage area.

[0111] When the task scheduler triggers an inference execution event, the probe injection module collects and records the original AE5 of the raw data fingerprint that the inference task needs to process by calling the target interface. Then, the normalization module preprocesses the original AE5 to obtain the normalized AE5, and the association module performs a hash operation on the normalized AE5 to obtain the log feature information SHA256(E5) of the node. SHA256(E5) represents the log feature information associated with the inference execution node E5, and the log feature information SHA256(E5) and the log information associated with the inference execution node E5 are stored in the storage area.

[0112] When the task scheduler triggers an output delivery event, the probe injection module collects the original AE6 of the log information recording the result data fingerprint of the inference task by calling the target interface. Then, the normalization module preprocesses the original AE6 to obtain the normalized AE6, and the association module performs a hash operation on the normalized AE6 to obtain the log feature information SHA256(E6) of the node. SHA256(E6) represents the log feature information associated with the output delivery node E6, and the log feature information SHA256(E6) and the log information associated with the output delivery node E6 are stored in the storage area.

[0113] When the task scheduler triggers a deterministic event, the probe injection module collects the raw AE7 of the application status information recorded in the result data by calling the target interface. Then, the normalization module preprocesses the raw AE7 to obtain a normalized AE7, and the association module performs a hash operation on the normalized AE7 to obtain the log feature information SHA256(E7) of the node. SHA256(E7) represents the log feature information associated with the determined node E7, and the log feature information SHA256(E7) and the log information associated with the determined node E7 are stored in the storage area.

[0114] The following is combined with Figure 4 The operation of storing log information in association according to logical relationships in the embodiments of this disclosure will be described in detail.

[0115] According to embodiments of this disclosure, the multiple lifecycle nodes include M, where M is an integer greater than 1; the log information is stored in association according to logical relationships, which may include the following operations: in response to receiving the log information of the m-th lifecycle node, reading the log feature information of the first m-1 lifecycle nodes from the storage area; m is an integer greater than 1 and less than or equal to M; performing feature extraction on the log information of the m-th lifecycle node to obtain the log feature information of the m-th lifecycle node; and concatenating the log feature information of the m-th lifecycle node with the log feature information of the first m-1 lifecycle nodes and writing it into the storage area.

[0116] Figure 4 This illustration shows a diagram illustrating the logical association and storage of log information from multiple lifecycle nodes according to an embodiment of the present disclosure.

[0117] like Figure 4 As shown, the lifecycle nodes of this example inference task can include five nodes, namely nodes E1 to E5. The log information of each node is obtained by calling the target node, namely logs AE1 to AE5.

[0118] Since the acquisition and storage of log information are carried out during the execution of the inference task, there are time-series differences in the log information generated by each node.

[0119] After hashing log AE1, the log characteristic information feature F1 of node E1 is obtained. Feature F1 is then stored in the storage area as the first event hash value of this inference task.

[0120] When log AE2 is received from node E2, it is first hashed to obtain the log feature information feature F2 of node E2. Then, the first event hash value "feature F1" of the inference task is read from the storage area, and feature F1 is concatenated with feature F2 and written to the storage area. In this way, the log information generated by all lifecycle nodes during the execution of the inference task is stored in a logical chain.

[0121] In some embodiments, the concatenated features can be hashed again, and the processed hash value can be written to the storage area. Besides secure hash algorithms, other cryptographic algorithms can also be used to extract features from log information; this disclosure does not limit the specific algorithms used.

[0122] For each newly received log message, after feature extraction, it is written to the storage area as an extension event of the logical chain in an append-only manner. This realizes a logical chain that meets the cryptographic anti-tampering requirements for the execution path information of the inference task. Any modification operation on the log message will be completed by appending the record, which further improves the security and anti-tampering difficulty of the log message.

[0123] Since the embodiments of this disclosure record multiple dimensions of information during the execution of the inference task, multi-dimensional information queries can be supported during the audit and traceability process of the inference task execution, such as queries of any one or more of the following dimensions: task identifier, time range, user dimension, model version dimension, etc.

[0124] The following is combined with Figure 5 The embodiments of this disclosure that perform auditing and tracing based on log information associated with each lifecycle node are described in detail.

[0125] According to embodiments of this disclosure, the above method may further include the following operations: sorting each log information according to the time when each log information is written to the storage area to obtain a log information sequence; processing each log information in the log information sequence to obtain log feature information of each log information; and in response to determining that the log feature information of each log information is the same as the log feature information already stored in the storage area, analyzing the log information sequence using a large model to generate execution path information for the inference task.

[0126] Figure 5 The illustration shows a schematic diagram of log querying and traceability analysis for inference tasks according to embodiments of the present disclosure.

[0127] like Figure 5 As shown, the requesting end can refer to the client where the relevant auditor is located. When the retrieval module receives a query request from the requesting end, it can read the retrieval log information from the storage area according to the query dimensions included in the query request and return the log information sequence.

[0128] In this embodiment of the disclosure, the log information sequence is obtained by sorting the log information of each lifecycle node of the inference task according to the time when each log information is written to the storage area.

[0129] The log information sequence uses the physical append order of the log information written to the storage area, rather than the simple temporal order of the log information generation. This is because, in this embodiment, the log information of each lifecycle node is automatically acquired by the probe injection module when each lifecycle event is triggered. However, the timestamps of probes from different lifecycle nodes may have slight deviations due to network latency or clock drift. Although timing correction is performed during the preprocessing stage, residual errors may still exist. The physical append order, monotonically increasing by the storage layer, ensures that the log information sequence verification is strictly consistent with the order in which the log information is written to the storage area, laying the foundation for integrity verification.

[0130] Then, each log message in the log message sequence can be processed to obtain the log feature information of each log message. For example, the same feature extraction algorithm used when writing log messages to the storage area can be used to extract the log feature information of each log message.

[0131] Next, the log information of each lifecycle node is verified point by point. Since each log feature information stored in the storage area is associated with the log feature information of the preceding node, when it is determined that the log feature information of each log information is the same as the log feature information already stored in the storage area, it can be determined that the log information sequence read from the storage area is consistent with the log information written to the storage area and has not been tampered with, thus meeting the authenticity requirements of audit analysis.

[0132] Finally, large models can be used to analyze the log information sequence to generate execution path information for inference tasks.

[0133] For example, in order to reduce the probability of large model hallucinations, the statements generated by the large model can be constrained from the prompts used to input the large model to be based on the log information. Alternatively, the large model can be restricted to output the execution path information of the inference task in a three-part structure of "execution summary - event description - data integrity statement".

[0134] Since large models are prone to generating information that does not actually exist in the log information sequence when generating long text, this embodiment of the disclosure can introduce a phantom reference verification mechanism to compare the log information referenced by the information output by the large model with the original log information sequence. When the log information referenced by the information output by the large model all exist in the original log information sequence, it is determined that the confidence level of the information output by the large model at this time meets the audit requirements.

[0135] For example, the audit evidence package output by the large model may include three layers of information, such as: information on the integrity verification of the log information sequence at the metadata level; a structured log information sequence at the event sequence level, including at least the initial data fingerprint, the result data fingerprint, resource allocation records, and user roles; and information at the reporting level describing the complete execution path of the inference task from submission to downstream confirmation in natural language.

[0136] By verifying the integrity of log information sequences, the information integrity requirements of audit scenarios are met. By using a large model, structured log information is transformed into execution path information that meets audit requirements, further improving information processing efficiency.

[0137] Figure 6 A block diagram of an information processing apparatus according to an embodiment of the present disclosure is shown schematically.

[0138] like Figure 6 As shown, the information processing device 600 may include an acquisition module 610 and a storage module 620.

[0139] The acquisition module 610 is used to obtain log information associated with multiple lifecycle nodes by calling the target interface during the execution of the inference task; wherein, the multiple lifecycle nodes indicate the logical relationship between multiple operations in the execution process of the inference task.

[0140] The storage module 620 is used to associate and store log information according to logical relationships, so as to trace the execution path of the reasoning task based on the log information.

[0141] According to embodiments of this disclosure, multiple lifecycle nodes include: a task submission node. Log information includes first log information. The acquisition module 610 may include a first acquisition submodule, configured to, in response to receiving a submission request for an inference task, acquire the first log information associated with the task submission node by calling a target interface; wherein the first log information indicates information about the inference task.

[0142] According to embodiments of this disclosure, the multiple lifecycle nodes include: an inference execution node. Log information includes second log information associated with the inference execution node. The acquisition module 610 may include a second acquisition submodule, a first extraction submodule, and a first writing submodule.

[0143] The second acquisition submodule is used to obtain the first initial log information associated with the inference execution node by calling the target interface.

[0144] The first extraction submodule is used to extract features from the initial data that the inference task needs to process in the memory of the execution process of the inference task, and obtain the initial data feature information.

[0145] The first writing submodule is used to write the initial data feature information into the first initial log information to obtain the second log information.

[0146] According to embodiments of this disclosure, multiple lifecycle nodes include: an output delivery node; log information includes third log information associated with the output delivery node. The acquisition module 610 may include a third acquisition submodule, a second extraction submodule, and a second writing submodule.

[0147] The third acquisition submodule is used to obtain the second initial log information associated with the output delivery node by calling the target interface.

[0148] The second extraction submodule is used to extract features from the result data of the inference task in the memory of the execution process of the inference task, and obtain the feature information of the result data.

[0149] The second writing submodule is used to write the result data feature information into the second initial log information to obtain the third log information.

[0150] According to embodiments of this disclosure, multiple lifecycle nodes include task queuing nodes; log information includes fourth log information associated with the task queuing nodes. The acquisition module 610 may include a fourth acquisition submodule, configured to acquire the fourth log information by calling a target interface in response to determining that the inference task is in a waiting state; wherein the fourth log information indicates the waiting state information of the inference task.

[0151] According to embodiments of this disclosure, multiple generation cycle nodes include resource allocation nodes; log information includes fifth log information associated with the resource allocation nodes. The acquisition module 610 may include a fifth acquisition submodule, configured to acquire the fifth log information by calling a target interface in response to determining that the inference task has been assigned to a target processing node; wherein the fifth log information indicates the resource allocation status information of the inference task.

[0152] According to embodiments of this disclosure, multiple lifecycle nodes include: a model loading node; and log information including sixth log information associated with the model loading node. The acquisition module 610 may include a sixth acquisition submodule, configured to acquire the sixth log information by calling a target interface in response to determining that the model used to perform the inference task is in a loaded state; wherein the sixth log information indicates the model's loading information.

[0153] According to embodiments of this disclosure, multiple lifecycle nodes include: a determination node; and log information including seventh log information associated with the determination node. The acquisition module 610 may include a seventh acquisition submodule, configured to, in response to receiving a determination operation for result data of the inference task, acquire the seventh log information by calling a target interface; wherein the seventh log information indicates the application status information of the result data.

[0154] According to embodiments of this disclosure, the multiple lifecycle nodes include M, where M is an integer greater than 1; the storage module may include a reading submodule, a feature extraction submodule, and a splicing submodule.

[0155] The read submodule is used to read the log feature information of the first m-1 lifecycle nodes from the storage area in response to receiving the log information of the m-th lifecycle node; m is an integer greater than 1 and less than or equal to M.

[0156] The feature extraction submodule is used to extract features from the log information of the m-th lifecycle node to obtain the log feature information of the m-th lifecycle node.

[0157] The splicing submodule is used to splice the log feature information of the m-th lifecycle node with the log feature information of the previous m-1 lifecycle nodes and then write it to the storage area.

[0158] According to embodiments of this disclosure, the information processing apparatus 600 may further include a sorting module, a processing module, and an analysis module.

[0159] The sorting module is used to sort the log information according to the time when each log information was written to the storage area, so as to obtain a log information sequence.

[0160] The processing module is used to process each log message in the log information sequence to obtain the log feature information of each log message.

[0161] The analysis module is used to analyze the log information sequence using a large model in response to the determination that the log feature information of each log message is the same as the log feature information already stored in the storage area, and to generate the execution path information of the inference task.

[0162] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0163] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods described above.

[0164] According to embodiments of the present disclosure, a non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions are used to cause a computer to perform the methods described above.

[0165] According to an embodiment of this disclosure, a computer program product includes a computer program that, when executed by a processor, implements the method described above.

[0166] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing an information processing method according to embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0167] like Figure 7 As shown, device 700 includes a computing unit 701, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 702 or a computer program loaded into random access memory (RAM) 703 from storage unit 708. The RAM 703 may also store various programs and data required for the operation of device 700. The computing unit 701, ROM 702, and RAM 703 are interconnected via bus 704. Input / output (I / O) interface 705 is also connected to bus 704.

[0168] Multiple components in device 700 are connected to I / O interface 705, including: input unit 706, such as keyboard, mouse, etc.; output unit 707, such as various types of monitors, speakers, etc.; storage unit 708, such as disk, optical disk, etc.; and communication unit 709, such as network card, modem, wireless transceiver, etc. Communication unit 709 allows device 700 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0169] The computing unit 701 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 701 performs the various methods and processes described above, such as information processing methods. For example, in some embodiments, the information processing method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 708. In some embodiments, part or all of the computer program may be loaded and / or installed on device 700 via ROM 702 and / or communication unit 709. When the computer program is loaded into RAM 703 and executed by the computing unit 701, one or more steps of the information processing method described above may be performed. Alternatively, in other embodiments, the computing unit 701 may be configured to perform information processing methods by any other suitable means (e.g., by means of firmware).

[0170] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0171] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0172] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0173] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0174] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0175] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0176] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0177] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An information processing method, comprising: During the execution of the inference task, log information associated with multiple lifecycle nodes is obtained by calling the target interface; wherein, the multiple lifecycle nodes indicate the logical relationships between multiple operations during the execution of the inference task; and The log information is associated and stored according to the logical relationship, so as to trace the execution path of the inference task based on the log information.

2. The method according to claim 1, wherein, The plurality of lifecycle nodes include: a task submission node; the log information includes first log information; During the execution of the inference task, the process of obtaining log information associated with multiple lifecycle nodes by calling the target interface includes: In response to receiving the submission request for the inference task, the first log information associated with the task submission node is obtained by calling the target interface; wherein the first log information indicates information about the inference task.

3. The method according to claim 1 or 2, wherein, The plurality of lifecycle nodes include: an inference execution node; the log information includes second log information associated with the inference execution node; During the execution of the inference task, the process of obtaining log information associated with multiple lifecycle nodes by calling the target interface includes: By calling the target interface, the first initial log information associated with the inference execution node is obtained; In the memory of the execution process of the inference task, feature extraction is performed on the initial data to be processed by the inference task to obtain initial data feature information; and The initial data feature information is written into the first initial log information to obtain the second log information.

4. The method according to any one of claims 1-3, wherein, The plurality of lifecycle nodes include: an output delivery node; the log information includes third log information associated with the output delivery node; During the execution of the inference task, the process of obtaining log information associated with multiple lifecycle nodes by calling the target interface includes: By calling the target interface, the second initial log information associated with the output delivery node is obtained; In the memory of the execution process of the inference task, feature extraction is performed on the result data of the inference task to obtain result data feature information; and The result data feature information is written into the second initial log information to obtain the third log information.

5. The method according to any one of claims 1-4, wherein, The plurality of lifecycle nodes include task queuing nodes; the log information includes fourth log information associated with the task queuing nodes; During the execution of the inference task, the process of obtaining log information associated with multiple lifecycle nodes by calling the target interface includes: In response to determining that the inference task is in a waiting state, the fourth log information is obtained by calling the target interface; wherein the fourth log information indicates the waiting state information of the inference task.

6. The method according to any one of claims 1-5, wherein, The plurality of generation cycle nodes include resource allocation nodes; the log information includes fifth log information associated with the resource allocation nodes; During the execution of the inference task, the process of obtaining log information associated with multiple lifecycle nodes by calling the target interface includes: In response to determining that the inference task has been assigned to the target processing node, the fifth log information is obtained by calling the target interface; wherein the fifth log information indicates the resource allocation status information of the inference task.

7. The method according to any one of claims 1-6, wherein, The multiple lifecycle nodes include: a model loading node; the log information includes sixth log information associated with the model loading node; During the execution of the inference task, the process of obtaining log information associated with multiple lifecycle nodes by calling the target interface includes: In response to determining that the model used to perform the inference task is in a loaded state, the sixth log information is obtained by calling the target interface; wherein the sixth log information indicates the loading information of the model.

8. The method according to any one of claims 1-7, wherein, The plurality of lifecycle nodes include: a determined node; the log information includes seventh log information associated with the determined node; During the execution of the inference task, the process of obtaining log information associated with multiple lifecycle nodes by calling the target interface includes: In response to receiving a determination operation for the result data of the inference task, the seventh log information is obtained by calling the target interface; wherein the seventh log information indicates the application status information of the result data.

9. The method according to any one of claims 1-8, wherein, The plurality of lifecycle nodes includes M nodes, where M is an integer greater than 1; The step of associating and storing the log information according to the logical relationship includes: In response to receiving log information from the m-th lifecycle node, read the log feature information of the previous m-1 lifecycle nodes from the storage area; m is an integer greater than 1 and less than or equal to M; Feature extraction is performed on the log information of the m-th lifecycle node to obtain the log feature information of the m-th lifecycle node; and The log feature information of the m-th lifecycle node is concatenated with the log feature information of the previous m-1 lifecycle nodes and then written into the storage area.

10. The method according to claim 9, further comprising: The log information is sorted according to the time when it is written to the storage area to obtain a log information sequence; The log information in the log information sequence is processed to obtain the log feature information of each log information; as well as In response to determining that the log feature information of each log message is the same as the log feature information already stored in the storage area, the log message sequence is analyzed using a large model to generate the execution path information of the inference task.

11. An information processing apparatus, comprising: The acquisition module is used to acquire log information associated with multiple lifecycle nodes by calling the target interface during the execution of the inference task; wherein, the multiple lifecycle nodes indicate the logical relationship between multiple operations in the execution process of the inference task. The storage module is used to associate and store the log information according to the logical relationship, so as to trace the execution path of the inference task based on the log information.

12. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-10.

13. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-10.

14. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-10.

Images