Information processing device, information processing method, and program.

Abstract

In a distributed inference environment, when external inputs change or the inference state is updated, the recalculation scope is appropriately localized while ensuring the consistency of the inference. [Solution] The inference state group is obtained as a high-dimensional state representation including feature vectors, constraint satisfaction information, inference dependency structure, and virtual node identifier VID. A filtration is constructed based on a distance scale and neighborhood rules or threshold sequences, and a topological analysis (including persistent homology analysis) is performed or incrementally updated. Column identifiers such as column IDs (single IDs) that were involved in determining the birth, disappearance, or change of topological features are extracted as trace tokens. The column identifiers contained in the trace tokens are mapped to single identifiers, and then mapped to the VID set via vertices to calculate the contribution. Based on the contribution, the influential VID set is determined, and a local recalculation is instructed to the computation node corresponding to that set.

Description

[Technical Field]

[0001] This disclosure relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] For example, in distributed inference environments, model parallelization techniques are known to improve the efficiency of training or inference processing by distributing large-scale language models across multiple computing nodes (see Non-Patent Literature 1). The model parallelization technique shown in Non-Patent Literature 1 discloses a configuration in which the matrix operations of the Transformer's attention head and multilayer perceptron blocks are divided and distributed across multiple GPUs and synchronized by a small number of communication primitives. A similar distributed deployment concept can also be applied to inference environments, realizing an environment in which inference models are distributed across multiple computing nodes. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, Bryan Catanzaro, “Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism”, arXiv preprint arXiv:1909.08053, 2019, [Retrieved May 27, 2020], Internet<URL : https: / / arxiv.org / abs / 1909.08053> [Non-Patent Document 2] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N.Gomez, Lukasz Kaiser, Illia Polosukhin, “Attention Is All You Need”, Advances in Neural Information Processing Systems 30, 2017, [Retrieved May 27, 2020], Internet<URL : https: / / arxiv.org / pdf / 1706.03762> [Non-Patent Document 3] Hugging Face, “Model outputs”, Transformers Documentation, v5.3.0, [Retrieved May 27, 2026], Internet<URL : https: / / huggingface.co / docs / transformers / main_classes / output> [Non-Patent Document 4] Hugging Face, “KV cache strategies”, Transformers Documentation, v5.3.0, [Retrieved May 27, 2026], Internet<URL : https: / / huggingface.co / docs / transformers / kv_cache> [Non-Patent Document 5] Hugging Face, “Caching”, Transformers Documentation, v5.3.0, [Retrieved May 27, 2026], Internet<URL : https: / / huggingface.co / docs / transformers / cache_explanation> [Non-Patent Document 6] Herbert Edelsbrunner, David Letscher, Afra Zomorodian, “Topological Persistence and Simplification”, Discrete & Computational Geometry, Vol.28, No.4, pp.511-533, 2002, [Retrieved May 27, 2026], Internet<URL : https: / / doi.org / 10.1007 / s00454-002-2885-2> [Non-Patent Document 7] Afra Zomorodian, Gunnar Carlsson, “Computing Persistent Homology”, Discrete & Computational Geometry, Vol.33, No.2, pp.249-274, 2005, [Retrieved May 27, 2026], Internet<URL : https: / / doi.org / 10.1007 / s00454-004-1146-y> [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] In such an inference environment, changes in external inputs, knowledge updates, or updates to some inference states may necessitate recalculations to ensure the consistency of the entire inference. Traditionally, the most direct method for ensuring such consistency has been global recalculation, which performs recalculations across the entire set of inference states. However, global recalculations involve instructing all distributed computing nodes to recalculate and aggregating the results, which can lead to excessive consumption of computing resources, strain on communication bandwidth, and increased response delays.

[0005] This disclosure is made to solve the above-mentioned problems and aims to provide an information processing device, an information processing method, and a program that can, in an inference environment where inference processing is performed using an inference model located on computing resources, identify with high accuracy the range of influence that changes in external inputs have on the inference state group, and instruct local recalculations limited to that range of influence. [Means for solving the problem]

[0006] One of the information processing devices disclosed herein controls an inference environment that performs inference processing using an inference model located on computing resources, (A) A state acquisition unit that acquires the inference state generated during the progress of the inference process as a high-dimensional state representation including a virtual node identifier VID defined on a virtual logical address space, (B) A phase analysis unit that performs or incrementally updates a filtration-based phase analysis on the set of inference states to calculate a topological feature that reflects at least one of the topological properties of the latent space or the structural changes of the computation graph, and generates correspondence information between the inference states and individual constituent vertices corresponding to the inference states in accordance with the configuration of the filtration, (C) Difference determination unit, triggered by a change in external input to at least a part of the inference state or an update of the inference state, extracts a record unit (hereinafter referred to as a trace token) from the boundary matrix reduction process in the internal processing of the phase analysis, which includes the identifier of the principal column generated in the process of the reduction process and the identifier of the erased column, which is a column added for the erasure process to the column to be reduced in the process of the reduction process, (D) A difference determination unit that maps the set of columns identified by the identifier of the source column and the identifier of the erased column contained in the trace token to a single identifier, obtains the set of vertices corresponding to the single identifier, further maps it to a set of VIDs based on the mapping information that references the virtual node identifier VID associated with the vertex, and identifies the set of influential VIDs that contributed to the local change on the topology based on the set of VIDs, (E) A recalculation control unit that issues a local recalculation command to the computing resource corresponding to the influencing VID set, including constraints and a group of VID identifiers to be recalculated, and causes the recalculation to be performed in a local range including the influencing VID set, It is equipped with.

[0007] One of the information processing methods disclosed herein is a method for controlling an inference environment that performs inference processing using an inference model located on computing resources, (A) A state acquisition step of acquiring the inference state generated during the progress of the inference process as a high-dimensional state representation including a virtual node identifier VID defined on a virtual logical address space, (B) A phase analysis step which involves performing or incrementally updating a filtration-based phase analysis on the set of inference states to calculate a topological feature that reflects at least one of the topological properties of the latent space or the structural changes of the computation graph, and generating correspondence information between the inference states and the individual constituent vertices corresponding to the inference states in accordance with the configuration of the filtration, (C) A difference determination step, triggered by a change in external input to at least a part of the inference state or an update of the inference state, extracts a record unit (hereinafter referred to as a trace token) from the boundary matrix reduction process in the internal processing of the phase analysis, which includes the identifier of the principal column generated in the process of the reduction process and the identifier of the erased column, which is a column added for the erasure process to the column to be reduced in the process of the reduction process. (D) A difference determination step which involves mapping the set of columns identified by the identifier of the source column and the identifier of the erased column contained in the trace token to a single identifier, obtaining the set of vertices corresponding to the single identifier, further mapping it to a set of VIDs based on the mapping information which refers to the virtual node identifier VID associated with the vertex, and identifying the set of influential VIDs that contributed to the local change on the topology based on the set of VIDs, (E) A recalculation control step which issues a local recalculation command to the computing resource corresponding to the influencing VID set, including constraints and a group of VID identifiers to be recalculated, and causes the recalculation to be performed corresponding to the local range including the influencing VID set, Includes.

[0008] One of the programs included in this disclosure is intended to cause a computer to execute the information processing methods of this disclosure. [Effects of the Invention]

[0009] According to the information processing device, information processing method, and program of this disclosure, in an inference environment in which inference processing is performed using an inference model located on computing resources, it is possible to identify with high accuracy the range of influence that a change in external input has on the group of inference states, and to instruct local recalculations limited to that range of influence. [Brief explanation of the drawing]

[0010] [Figure 1] This block diagram shows a schematic configuration of an information processing device and a distributed inference environment according to one embodiment of the present disclosure. [Figure 2] This is an explanatory diagram illustrating the general process of acquiring the inference state, configuring the filtration, and performing phase analysis (including persistent homology analysis). [Figure 3] This flowchart shows the processing flow for identifying column identifiers, resolving the mapping (column identifier → individual → vertex → VID), calculating contribution, and determining the set of influential VIDs. [Figure 4] This flowchart outlines the general flow of the process for instructing local recalculations, calculating SavedCost, and converting and finalizing the number of search branches or beam width. [Figure 5] This diagram outlines the safety transition, analysis resolution switching, hysteresis application, and (optional) backup / restart package processing based on error boundary information and reliability information. [Figure 6] This diagram illustrates an example of consistency auditing based on topological features, showing persistence diagrams (0-dimensional and 1-dimensional) calculated for a group of structured states (Consistent) and a group of unstructured states (Broken), as well as an overview of the calculation and determination (thresholding or hysteresis determination) of a consistency score (ConsScore) based on the duration of 1-dimensional features. [Figure 7] This diagram outlines the process of extracting VIDs from column identifiers (primary column, deletion column, or column set based on the origin set) via individual elements and vertices, determining the influential VID set by aggregating contributions using duration weighting (sequential correspondence between column identifiers, individual elements, vertices, VIDs, contributions, and influential VID set), mapping resolution limited to generation tag matching, and safe transitions in case of mismatch (extension to wide-area recalculation). [Figure 8] This diagram outlines the construction of a simplicial complex using a conversation log, column reduction processing, and recording of trace tokens according to Example 1, where (a) shows the simplicial complex and filtration configuration (σ1~σ13) of 5 vertices (v1~v5), (b) shows the recording of the deleted column in the reduction processing of columns σ9 and σ11, as well as the confirmation of the extinction of one-dimensional topological features by columns σ12 and σ13 and the calculation of the duration length of birth / extinction pairs, and (c) shows the construction of trace tokens A (duration length 1) and B (duration length 2) and the addition of duration length weighting contributions from the constituent vertices to the VID. [Figure 9] This is an explanatory diagram illustrating the aggregated contribution results for each VID and the identification of the influential VID set in Example 1. It shows the breakdown of contributions from trace tokens A (duration length 1) and B (duration length 2), indicating that VID(v3) has the highest contribution (=3) as a hub involved in both topological features, and that the influential VID set {VID(v3), VID(v4), VID(v5)} is identified by a contribution threshold ≥ 2. [Modes for carrying out the invention]

[0011] <Technical fields of this disclosure> This disclosure relates to an inference governance control device, method, and program for identifying the scope of impact and performing local recalculations and resource conversion and allocation based on persistent homology analysis. Specifically, it relates to a governance control technique that evaluates the set of inference states generated in a distributed inference environment using persistent homology analysis, identifies the scope of impact due to changes in external inputs etc. on a per-virtual node identifier (VID) basis, performs local recalculations, and further converts and allocates the surplus resources corresponding to the avoided recalculations to the number of search branches or beam width.

[0012] <Prior technologies and challenges> (Existing techniques for recalculation in distributed inference environments) As mentioned in the background technology section, model parallelization techniques are known for distributing large-scale language models across multiple computing nodes to improve the efficiency of training or inference processing in distributed inference environments (see, for example, Non-Patent Document 1). However, in such distributed inference environments, if external inputs are changed, knowledge is updated, or some inference states are updated, recalculation may be necessary to ensure the consistency of the entire inference. Conventionally, the most direct method for ensuring such consistency is global recalculation, which performs recalculation across the entire set of inference states. However, global recalculation involves instructing all distributed computing nodes to recalculate and aggregating the results, which can lead to excessive consumption of computing resources, strain on communication bandwidth, and increased response delays.

[0013] On the other hand, in autoregressive generation, a KV cache configuration is widely used that utilizes the causal attention property to cache and reuse key-value pairs corresponding to already processed tokens (see, for example, Non-Patent Documents 4 and 5). This KV cache, based on the property that the representation of past tokens remains invariant for subsequent tokens due to the causal mask, calculates attention at each generation step based on the query vector of the current token and the past keys and values ​​stored in the cache, thereby linearly suppressing the attention calculation cost per step with respect to the sequence length. Furthermore, implementations of the KV cache are known, including configurations that dynamically expand the cache size, configurations that pre-allocate a fixed-length cache, and configurations that reduce memory usage by quantizing the KV value with low precision (see, for example, Non-Patent Document 4). In addition, in publicly available implementations, configurations are known that allow the acquisition or retrieval of hidden states, attention weights, and past key-value pairs as the output of the inference model (see, for example, Non-Patent Document 3).

[0014] However, the KV cache configurations described in these non-patent documents are intended to improve the efficiency of cache accumulation and reuse, and do not disclose a configuration that quantitatively identifies the scope of the impact of changes in external inputs on the inference state group and instructs selective invalidation or local recalculation of the cache limited to that scope of impact.

[0015] Furthermore, regarding the identification of the scope of influence, there are known configurations that use graph database technology or graph search technology to extract a range reachable from a specific node or a group of related nodes that satisfy predetermined conditions on a relational graph consisting of predefined nodes and edges. However, these technologies assume that the relationships between objects are explicitly described in advance as nodes, edges, labels, or attributes. For this reason, no configuration has been disclosed that identifies the scope of influence via trace tokens derived from the reduction of boundary matrices for objects where relationships are not given in advance as an explicit graph, such as the inference state of a neural network, but appear indirectly as neighborhood relationships, connection relationships, or changes in the topological structure on a high-dimensional latent space.

[0016] (Applications of attention mechanisms and topological methods) In Transformer-based models, attention mechanisms based on Scaled Dot-Product Attention are widely used (see, for example, Non-Patent Document 2). This attention mechanism calculates weighted connections between input tokens based on query (Q), key (K), and value (V) matrices projected from each position in the input sequence, enabling the capture of dependencies between arbitrary positions within the sequence. Furthermore, multi-head attention allows multiple attention heads to learn dependencies on different subspaces in parallel, and these outputs are integrated to generate an output representation. While this attention mechanism contributes significantly to representing relationships between inputs in inference processing, it does not provide a means to monitor the structural stability of the resulting inference state group during inference execution, nor does it provide a means to control the scope of recalculation based on changes in that structure.

[0017] On the other hand, in the field of topology, mathematical methods have been established to perform filtration-based persistent homology analysis on point cloud data and other data to calculate topological features (persistence diagrams, Betti numbers, durations, etc.) (see, for example, Non-Patent Documents 6 and 7). Persistent homology analysis tracks homology groups on a simplicial complex that grow stepwise in response to changes in parameters (distance thresholds, etc.), and records the birth and death of topological features (connected components, loops, gaps, etc.). In the algorithm of this analysis, a reduction process (column operation) is performed on the boundary matrix, and the birth and death columns of each topological feature are identified by this reduction process. Persistent homology analysis is applied as a topological descriptor to grasp the global structure of data in fields such as data analysis, shape recognition, materials science, and sensor networks. However, there is no known configuration that extracts a group of identifiers of columns involved in the reduction process from the intermediate process of the boundary matrix reduction process and repurposes this group of identifiers for identifying the scope of influence of the inference state or for governance control.

[0018] (Differences between the technology / concepts presented in prior art documents and the scope of this disclosure) As described above, the attention mechanism shown in Non-Patent Document 2, and the hidden state / KV cache acquisition / reuse configurations shown in Non-Patent Documents 3-5 are technologies related to the execution and efficiency of inference processing. Model parallelization shown in Non-Patent Document 1 is a technology that realizes the physical distributed arrangement of inference models. Furthermore, persistent homology analysis shown in Non-Patent Documents 6 and 7 is a technology that establishes methods and algorithms for calculating topological features based on filtration.

[0019] While these prior art documents disclose useful configurations in their respective technical domains, none of them disclose or suggest a series of governance control configurations that include: (i) applying persistent homology analysis to a set of inference states in a distributed inference environment to calculate topological features; (ii) extracting recording units, including identifiers for the principal and elimination columns, from the intermediate process of boundary matrix reduction in the said analysis, and identifying the topological influence range via a virtual node identifier (VID) associated with a single constituent vertex; and (iii) allocating computational resources equivalent to recalculations avoided by local recalculations limited to the said influence range by converting them into the number of search branches or beam width within the same inference task.

[0020] As described above, a challenge in conventional distributed inference environments is that global recalculations resulting from some state updates lead to excessive consumption of computing resources, poor power efficiency, and increased response delays. On the other hand, existing local recalculation methods lack objective indicators for accurately defining the complex logical dependencies and causal influence scope between inference states in high-dimensional latent spaces. Therefore, there are technical challenges in avoiding inconsistencies in inference results due to insufficient recalculations, or unnecessary recalculations due to overestimation of the influence scope, resulting in inefficient resource utilization.

[0021] Another aspect of this disclosure is the need for a framework to ensure consistency in the dynamic reallocation of computing resources, such as limiting to incomplete branches or setting allocation conditions, in order to avoid misallocation and system instability when enhancing exploration within the same inference task by effectively utilizing savedcost recalculation resources avoided by local recalculation.

[0022] Another challenge that this disclosure aims to address is achieving both computational load reduction and accuracy in persistent homology analysis. When introducing techniques such as approximation, summarization, or distributed updating, it is desirable to establish safe transitions to wide-area recalculations based on error boundaries, or to establish a phased operation that switches to high-resolution estimation only when uncertainty increases, in order to address the numerical errors that may arise.

[0023] In a distributed environment, duplicate receipt of allocation confirmation messages and restart packages may occur due to network delays and retransmission processes. This device incorporates a version control function using idempotent keys and version numbers assigned to each message, thereby physically preventing the duplicate application of the same processing request and ensuring consistency of the inference state and accuracy of resource allocation.

[0024] In the boundary matrix reduction process, the mapping between column identifiers (original column and deletion column) and virtual node identifiers (VIDs) is dynamically managed based on computation generation tags, etc. Therefore, to eliminate the possibility of misidentifying IDs between different computation generations, a consistency verification procedure is provided that compares the generation tags contained in the trace token with the mapping information, thereby technically guaranteeing the accuracy of identifying the scope of influence.

[0025] The inference dependency structure may not be fully observed or consistent immediately after an update due to distributed update delays, missing data, or partial unobtained references. In this case, estimating the scope of influence based solely on the dependency structure's closure rule may result in omissions or overestimations. On the other hand, topological features based on changes in distances and neighborhood relationships between high-dimensional state representations of the inference state group can reflect signs of change even if the dependency structure is incomplete, thus contributing to the identification of the scope of influence.

[0026] The problems that each technology included in this disclosure aims to solve are summarized below. However, each technology included in this disclosure does not solve all of the following problems 1-6; the problems that can be solved vary depending on the elements included in each technology. Furthermore, each technology included in this disclosure may solve problems other than those listed below. These problems will be clarified in the descriptions of the elements included in each technology.

[0027] (Task 1) Capture changes in the inference state group as changes in topological features, identify the influencing VID set by calculating the contribution using column identifiers in the boundary matrix reduction process, and appropriately limit the scope of local recalculation.

[0028] (Challenge 2) Quantify the saved cost, which corresponds to recalculations avoided by local recalculations, and convert and allocate it to incomplete branches belonging to the search frontier within the same inference task as the number of search branches or beam width to improve search efficiency.

[0029] (Problem 3) Suppress misallocation and instability while reducing the computational load of phase analysis by using safe transitions based on error boundary information and reliability information, stepwise switching of analysis resolution, and stabilization by hysteresis.

[0030] (Challenge 4) Enable the continuation of updating the phase analysis through missing data imputation estimation even when missing data exists, and reflect the uncertainty caused by imputation in the error boundary information and reliability information.

[0031] (Challenge 5) To ensure operational continuity (fallback) in situations where the identification of the core influential VID set and local recalculation may fail, by using optional functions such as backup / notification / restart packages and ReleaseCost cycles, while suppressing duplicate application through version control and idempotency.

[0032] (Challenge 6) Ensure feasibility and auditability by tracking column identifiers in boundary matrices and clarifying column correspondence during incremental updates (assigning column IDs, tracing origins, managing generations, etc.).

[0033] <Technologies included in this disclosure> Each of the following [1] to

[39] is an example of the technology included in this disclosure. However, these are merely examples, and various other technologies described later are also included.

[0034] [1] An information processing device that controls an inference environment that performs inference processing using an inference model located on computing resources, (A) A state acquisition unit that acquires the inference state generated during the progress of the inference process as a high-dimensional state representation including a virtual node identifier VID defined on a virtual logical address space, (B) A phase analysis unit that performs or incrementally updates a filtration-based phase analysis on the set of inference states to calculate a topological feature that reflects at least one of the topological properties of the latent space or the structural changes of the computation graph, and generates correspondence information between the inference states and individual constituent vertices corresponding to the inference states in accordance with the configuration of the filtration, (C) Difference determination unit, triggered by a change in external input to at least a part of the inference state or an update of the inference state, extracts a record unit (hereinafter referred to as a trace token) from the boundary matrix reduction process in the internal processing of the phase analysis, which includes the identifier of the principal column generated in the process of the reduction process and the identifier of the erased column, which is a column added for the erasure process to the column to be reduced in the process of the reduction process, (D) A difference determination unit that maps the set of columns identified by the identifier of the source column and the identifier of the erased column contained in the trace token to a single identifier, obtains the set of vertices corresponding to the single identifier, further maps it to a set of VIDs based on the mapping information that references the virtual node identifier VID associated with the vertex, and identifies the set of influential VIDs that contributed to the local change on the topology based on the set of VIDs, (E) A recalculation control unit that issues a local recalculation command to the computing resource corresponding to the influencing VID set, including constraints and a group of VID identifiers to be recalculated, and causes the recalculation to be performed in a local range including the influencing VID set, An information processing device characterized by comprising:

[0035] [2] The information processing apparatus described in [1], Based on the determination of the influence VID set by the difference determination unit and the instruction for local recalculation by the recalculation control unit, The amount of physical computing resources required when performing a wide-area recalculation without using the aforementioned VID set, The difference between the amount of physical computing resources required when performing a local recalculation based on the aforementioned set of influential VIDs and the actual amount of resources required is A unit for calculating the remaining capacity, which is calculated as SavedCost, A resource conversion and allocation unit converts the SavedCost into a logical parameter (hereinafter referred to as ΔB) such as an additional number of search branches or an increment in beam width based on the resource / logical conversion coefficient η, and redistributes and applies the ΔB to parallel branches belonging to incomplete search frontiers within the same inference task. Furthermore, The information processing device is characterized in that the SavedCost is quantified as at least one of the number of floating-point operations reduced by avoiding recalculations, memory bandwidth usage, communication bandwidth usage, or the reduction in execution time.

[0036] [3] In the information processing device described in [1] or [2], The information processing apparatus is characterized in that the state acquisition unit acquires the high-dimensional state representation, which includes spatiotemporal coordinates that map the temporal delay associated with the inference process as a physical cost.

[0037] [4] In the information processing device described in any of [1] to [3], The information processing device is characterized in that the phase analysis unit derives a phase entropy indicating the structural stability of inference from the phase features.

[0038] [5] In the information processing device described in any of [1] to [4], The difference determination unit is characterized by calculating the contribution of each VID by weighting the VID set based on the duration of the topological feature, and identifying the influencing VID set based on said contribution.

[0039] In the information processing apparatus described in any of [1] to [5], The information processing device is characterized in that the local recalculation instruction includes a calculation generation tag for the recalculation.

[0040] [7] In an information processing device described in any of [1] to [6], The information processing device is characterized in that the constraint conditions include a constraint condition that fixes the state value of the boundary VID located at the outer edge of the set of influential VIDs as a constant.

[0041] In the information processing device described in any of [1] to [7], The difference determination unit performs pruning to exclude from the calculation of contribution column identifiers among the column identifiers included in the trace token whose duration of the corresponding topological feature is less than a predetermined noise threshold, and identifies the column identifier by including, in addition to the identifier of the principal column and the identifier of the deleted column, a set of column IDs included in the origin set recorded by the topological analysis unit.

[0042] In the information processing device described in any of [9] [1] to [8], The aforementioned correspondence information includes a generation tag that matches the column ID and generation number, The difference determination unit, limited to references that share the same generation tag, resolves the mapping to the VID based on the mapping from the column identifier through the individual to the vertex. An information processing device characterized by, if the generation tags do not match, ceasing the identification of the influencing VID set and expanding the recalculation target to a predetermined wide range.

[0043]

[10] In the information processing device described in any of [1] to [9], The difference determination unit is characterized in that, when there is a many-to-many correspondence between the individual components and the VID, it aggregates or allocates contributions according to at least one allocation rule of equal distribution, aggregation by priority rule, or reliability weight distribution.

[0044]

[11] In the information processing device described in any of [1] to

[10] , The phase analysis unit calculates an upper limit or an estimated value of the error introduced due to at least one of summarization, approximation, distributed updating, or reduction of analysis resolution as error boundary information during the execution or incremental update of the phase analysis based on the filtration, and stores reliability information based on the error boundary information in association with the virtual node identifier VID. The difference determination unit, if the reliability information is less than a predetermined reliability threshold, stops identifying the influencing VID set and expands the recalculation target to a predetermined wide area. The information processing device is characterized in that the phase analysis unit attempts to fill in any missing information in the inference state group or inference dependency structure by retrospective analysis or neighborhood matching based on the inference dependency structure.

[0045] In the information processing apparatus described in

[12]

[11] , The difference determination unit or the recalculation control unit determines data health based on at least one of the reliability information, the error boundary information, the reliability score received from an external device, or the alternative reliability score calculated internally by the device, and suppresses chattering by applying hysteresis characteristics to at least one of the transition from local recalculation to global recalculation, return from global recalculation to local recalculation, or specific execution and cessation of the influencing VID set. An information processing device characterized in that, if the data integrity does not meet predetermined conditions or if the error boundary information increases by a predetermined amount or more, the application of resource conversion allocation is temporarily suspended, undetermined allocations are invalidated or the allocation amount is reduced, and the recalculation control unit transitions to wide-area recalculation to restore consistency of the inference state.

[0046] In the information processing apparatus described in

[13] [2], At least one of the above-level features, the contribution of each VID if calculated, the set of influential VIDs, the error boundary information or reliability information if calculated, or the history or parameters related to SavedCost is associated with the user identifier UID, the inference task identifier TID, or the integrated identifier UTID generated from the combination of the UID and TID. Based on the TID, UID, or UTID, phase transition profiles are accumulated in a time series to identify or update the inference signature. The information processing device is characterized in that the reference of the aforementioned inference signature is limited to governance applications including the identification of the set of influencing VIDs, the instruction of local recalculation, the allocation of resource conversions, and the selection or correction of thresholds or weights used therein.

[0047] In the information processing apparatus described in

[14] [2] or

[13] , The difference determination unit or the phase analysis unit records the identifier included in the trace token, the VID identification result based on the correspondence information, the contribution of each VID if calculated, and at least one of the influential VID set as an audit log. At least one of the recalculation control unit or the resource transfer and allocation unit generates an audit record relating to the execution history of the local recalculation or the transfer and allocation of resources. The aforementioned audit record includes a hash chain that contains the hash value of the previous audit record. The recalculation control unit or the resource transfer and allocation unit is configured to output the audit record to an external audit system after attaching an electronic signature or message authentication code to it, and the request or response sent and received with an external computing node or state storage device includes at least an idempotent key, and the duplicate application of the request is suppressed or suppressed by an external device when the request containing the same idempotent key is resent or received multiple times.

[0048]

[15] In the information processing device described in any of [1] to

[14] , The system further includes an indicator receiving unit that receives from an external device at least one of the following: an uncertainty indicator representing the integrity of the input, syntactic uncertainty, semantic uncertainty, or uncertainty of constraint satisfaction, and an action-related indicator related to the user's input generation process or dialogue operation. The difference determination unit, when the uncertainty index satisfies a predetermined safety condition, stops identifying the set of influencing VIDs and expands the recalculation target to a predetermined wide range. If the behavior-related indicator satisfies predetermined deterrent conditions, at least one of the threshold used to determine the safe condition and the distance threshold used to determine the change in the topological feature is corrected in the direction of safety. An information processing device characterized by suppressing chattering by applying a hysteresis characteristic that makes the suppression condition and the recovery condition different.

[0049]

[16] In the information processing device described in any of [1] to

[15] , The recalculation control unit generates a local control command for at least one of the inference state, intermediate representation, cache area, compute node, or communication path corresponding to the influencing VID set, based on the result of identifying the influencing VID set. The information processing apparatus is characterized in that the local control instruction includes at least one of the following: local recalculation, integrity verification, cache invalidation, state value fixing, output reflection temporary suspension, processing priority change, computing resource allocation change, communication bandwidth allocation change, state saving, audit identifier issuance, impact scope record generation, integrity verification hash calculation, or transition to wide-area recalculation.

[0050] In the information processing apparatus described in

[17] [2],

[13] or

[14] , The recalculation control unit further includes a transfer unit that transfers at least a portion of the intermediate inference results associated with the VID to an external state storage device that stores the inference state as static data if the recalculation control unit fails to perform a local recalculation or if the difference determination unit detects an abnormal state that makes recalculation impossible. The external state storage device, which is the transfer destination of the transfer unit, is characterized in that it generates and stores a restart anchor, which serves as the starting point for restarting the inference, and a structural identifier that identifies the unfulfilled conditions or scope of influence necessary for restarting, based on the transferred intermediate inference results.

[0051] In the information processing apparatus described in

[18]

[17] , The restart package, which includes the restart anchor and the structure identifier, includes an idempotent key, and based on this idempotent key, duplicate application during retransmission or duplicate receipt is suppressed. An information processing device characterized by defining the resources released by the transfer or evacuation as ReleasedCost, adding it to SavedCost, or separately summing them up and treating them as surplus, and applying them as an additional term to the number of search branches or beam width.

[0052] In the information processing apparatus described in

[19] [2],

[13] ,

[14] ,

[17] , or

[18] , The recalculation control unit, in conjunction with the result of the difference determination unit's identification of the influence VID set, stops or suppresses all or part of the subsequent inference or recalculation process for processing sections not belonging to the influence VID set, based on whether the nearest match rate or matching success rate within a predetermined time window satisfies predetermined conditions. The computing resources avoided by the shutdown or suppression are calculated and added to SavedCost, The SavedCost is preferentially reallocated to local recalculations of processing intervals belonging to the affected VID set. An information processing device characterized by the following:

[0053]

[20] An information processing method for controlling an inference environment that performs inference processing using an inference model located on computing resources, (A) A state acquisition step of acquiring the inference state generated during the progress of the inference process as a high-dimensional state representation including a virtual node identifier VID defined on a virtual logical address space, (B) A phase analysis step which involves performing or incrementally updating a filtration-based phase analysis on the set of inference states to calculate a topological feature that reflects at least one of the topological properties of the latent space or the structural changes of the computation graph, and generating correspondence information between the inference states and the individual constituent vertices corresponding to the inference states in accordance with the configuration of the filtration, (C) A difference determination step, triggered by a change in external input to at least a part of the inference state or an update of the inference state, extracts a record unit (hereinafter referred to as a trace token) from the boundary matrix reduction process in the internal processing of the phase analysis, which includes the identifier of the principal column generated in the process of the reduction process and the identifier of the erased column, which is a column added for the erasure process to the column to be reduced in the process of the reduction process. (D) A difference determination step which involves mapping the set of columns identified by the identifier of the source column and the identifier of the erased column contained in the trace token to a single identifier, obtaining the set of vertices corresponding to the single identifier, further mapping it to a set of VIDs based on the mapping information which refers to the virtual node identifier VID associated with the vertex, and identifying the set of influential VIDs that contributed to the local change on the topology based on the set of VIDs, (E) A recalculation control step which issues a local recalculation command to the computing resource corresponding to the influencing VID set, including constraints and a group of VID identifiers to be recalculated, and causes the recalculation to be performed corresponding to the local range including the influencing VID set, An information processing method characterized by including

[0054]

[21] In the information processing method described in

[20] , Based on the determination of the influencing VID set by the difference determination step and the instruction for local recalculation by the recalculation control step, the difference between the amount of physical computing resources required when performing a wide-area recalculation without using the influencing VID set and the amount of physical computing resources required when performing a local recalculation based on the influencing VID set is calculated as SavedCost. The process further includes converting the SavedCost into logical parameters such as the number of search branches or the increment of the beam width (hereinafter referred to as ΔB) based on the resource / logical conversion coefficient η, and reallocating the ΔB to parallel branches belonging to incomplete search frontiers within the same inference task. The information processing method is characterized in that the SavedCost is quantified as at least one of the number of floating-point operations reduced by avoiding recalculations, memory bandwidth usage, communication bandwidth usage, or the reduction in execution time.

[0055] In the information processing method described in

[22]

[20] or

[21] , The information processing method is characterized in that, in the state acquisition step, the high-dimensional state representation is acquired, which includes spatiotemporal coordinates that map the temporal delay associated with the inference process as a physical cost.

[0056] In the information processing method described in any of

[20] to

[22] , The information processing method is characterized in that, in the phase analysis step, a phase entropy indicating the structural stability of inference is derived from the phase features.

[0057] In the information processing method described in any of

[20] to

[23] , The information processing method is characterized in that, in the difference determination step, the VID set is weighted based on the duration of the topological features to calculate the contribution of each VID, and the influencing VID set is identified based on the contribution.

[0058] In the information processing method described in any of

[20] to

[24] , The information processing method is characterized in that the local recalculation instruction includes a calculation generation tag for the recalculation.

[0059] In the information processing method described in any of

[20] to

[25] , The information processing method is characterized in that the constraint conditions include a constraint condition that fixes the state value of the boundary VID located at the outer edge of the set of influential VIDs as a constant.

[0060] In the information processing method described in any of

[20] to

[26] , The information processing method is characterized in that, in the difference determination step, a pruning process is performed to exclude from the calculation of contribution column identifiers among the column identifiers included in the trace token whose duration of the corresponding topological feature is less than a predetermined noise threshold, and the column identifier is identified by including, in addition to the identifier of the principal column and the identifier of the deleted column, a set of column IDs included in the origin set recorded in the topological analysis step.

[0061] In the information processing method described in any of

[20] to

[27] , The aforementioned correspondence information includes a generation tag that matches the column ID and generation number, In the aforementioned difference determination process, the mapping to VID is resolved based on the mapping from the column identifier through the individual to the vertex, limited to references that share the same generation tag. An information processing method characterized by, if the generation tags do not match, ceasing the identification of the influencing VID set and expanding the recalculation target to a predetermined wide area.

[0062] In the information processing method described in any of

[20] to

[28] , The information processing method is characterized in that, in the difference determination step, when there is a many-to-many correspondence between the individual components and the VID, contributions are aggregated or allocated according to at least one allocation rule of equal distribution, aggregation by priority rule, or reliability weight distribution.

[0063] In the information processing method described in any of

[30]

[20] to

[29] , In the phase analysis step, an upper limit or an estimate of the upper limit of the error introduced due to at least one of summarization, approximation, distributed updating, or reduction of analysis resolution is calculated as error boundary information, and reliability information based on the error boundary information is stored in association with the virtual node identifier VID. In the difference determination step, if the reliability information is less than a predetermined reliability threshold, the identification of the influencing VID set is stopped and the recalculation target is expanded to a predetermined wide area. The information processing method is characterized in that, in the phase analysis step, if there are gaps in the inference state group or inference dependency structure, an attempt is made to fill them in by retrospection or neighborhood matching based on the inference dependency structure.

[0064]

[31] In the information processing method described in

[30] , In the difference determination step or the recalculation control step, data health is determined based on at least one of the reliability information, the error boundary information, the reliability score received from an external device, or the alternative reliability score calculated internally by the device, and hysteresis characteristics are applied to at least one of the transition from local recalculation to global recalculation, return from global recalculation to local recalculation, or specific execution and cessation of the influencing VID set to suppress chattering. An information processing method characterized by temporarily suspending the application of resource conversion allocation, invalidating undetermined allocations, or reducing the allocation amount when data integrity does not meet predetermined conditions or when the error boundary information increases by a predetermined amount or more, thereby transitioning to wide-area recalculation and restoring consistency of the inference state.

[0065] In the information processing method described in

[32]

[21] , At least one of the above-level features, the contribution of each VID if calculated, the set of influential VIDs, the error boundary information or reliability information if calculated, or the history or parameters related to SavedCost is associated with the user identifier UID, the inference task identifier TID, or the integrated identifier UTID generated from the combination of the UID and TID. An information processing method characterized by limiting the reference of the inference signature to governance purposes, including identifying the set of influencing VIDs, instructing local recalculations, allocating resource redirects, and selecting or correcting thresholds or weights used therein.

[0066] In the information processing method described in

[33]

[21] or

[32] , The identifier included in the trace token, the VID identification result based on the mapping information, the contribution of each VID if calculated, and at least one of the influential VID set are recorded as an audit log. Generate audit records relating to the execution history of the aforementioned local recalculations or the redirection and allocation of resources. The aforementioned audit record is given a hash chain that includes the hash value of the previous audit record. The aforementioned audit record is configured to be output to an external audit system after being given an electronic signature or message authentication code. An information processing method characterized by including at least an idempotent key in requests or responses transmitted to an external computing node or state storage device, and preventing the duplicate application of such requests, or causing the external device to prevent duplicate application, if such requests containing the same idempotent key are retransmitted or received multiple times.

[0067] In the information processing method described in any of

[34]

[20] to

[33] , The system receives from an external device an uncertainty index representing the integrity of the input, syntactic uncertainty, semantic uncertainty, or uncertainty of constraint satisfaction, and at least one of the behavior-related indexes related to the user's input generation process or dialogue operation. If the uncertainty index satisfies a predetermined safety condition, the identification of the influencing VID set is stopped and the recalculation target is expanded to a predetermined wide range. If the behavior-related indicator satisfies predetermined deterrent conditions, at least one of the threshold used to determine the safe condition and the distance threshold used to determine the change in the topological feature is corrected in the direction of safety. An information processing method characterized by suppressing chattering by applying a hysteresis characteristic that makes the suppression condition and the recovery condition different.

[0068] In the information processing method described in any of

[35]

[20] to

[34] , The recalculation control step generates a local control instruction for at least one of the inference state, intermediate representation, cache area, compute node, or communication path corresponding to the influencing VID set, based on the result of identifying the influencing VID set. The information processing method is characterized in that the local control instruction includes at least one of the following: local recalculation, integrity verification, cache invalidation, state value fixing, output reflection temporary suspension, processing priority change, computing resource allocation change, communication bandwidth allocation change, state saving, audit identifier issuance, impact scope record generation, integrity verification hash calculation, or transition to wide-area recalculation.

[0069] In the information processing method described in

[36]

[21] ,

[32] or

[33] , If local recalculation fails or an abnormal state that makes recalculation impossible is detected, at least a portion of the intermediate inference results associated with the VID is transferred to an external state storage device that stores the inference state as static data. An information processing method characterized in that the external state storage device, which is the transfer destination of the transfer unit, generates and stores a restart anchor that serves as the starting point for restarting the inference, and a structural identifier that identifies the unfulfilled conditions or scope of influence necessary for restarting, based on the transferred intermediate inference results.

[0070] In the information processing method described in

[37]

[36] , The restart package, which includes the restart anchor and the structure identifier, includes an idempotent key, and based on this idempotent key, duplicate application during retransmission or duplicate receipt is suppressed. An information processing method characterized by defining the resources released by the aforementioned transfer or evacuation as ReleasedCost, adding it to the SavedCost or separately summing them up and treating them as surplus, and applying them as an additional term to the number of search branches or beam width.

[0071] In the information processing methods described in

[38]

[21] ,

[32] ,

[33] ,

[36] or

[37] , In conjunction with the identification of the aforementioned influential VID set, for processing intervals not belonging to the aforementioned influential VID set, if the nearest neighbor match rate or matching success rate within a predetermined time window satisfies predetermined conditions, all or part of the subsequent inference processing or recalculation processing is stopped or suppressed. The computing resources avoided by the shutdown or suppression are calculated and added to SavedCost, An information processing method characterized by preferentially reallocating the SavedCost to the local recalculation of processing intervals belonging to the influence VID set.

[0072]

[39] A program for causing a computer to execute any of the information processing methods described in

[20] through

[38] .

[0073] <Key points of this disclosure> This disclosure obtains an inference state as a high-dimensional state representation and calculates topological features by performing or incrementally updating a sustained homology analysis based on filtration.

[0074] This disclosure provides a configuration that enables evaluation of the structural stability of the inference state group during the inference process by performing or incrementally updating a filtration-based persistent homology analysis on the inference state group generated by the attention mechanism of an architecture such as a Transformer to calculate topological features.

[0075] This disclosure includes a configuration that performs a phase analysis on a set of inference states, including hidden states, key-value pairs, or high-dimensional state representations based thereon, and, triggered by a change in external input or an update of the inference states, identifies the scope of influence on a unit of virtual node identifiers (VIDs) based on trace tokens extracted from the reduction process of the boundary matrix, and instructs a local recalculation limited to that scope of influence.

[0076] In this disclosure, the topological dependencies of a group of inference states distributed across multiple computing nodes are analyzed by persistent homology analysis, and the groups of simplices that contributed to the changes in topological features in the analysis results are identified as trace tokens, thereby determining the targets for local recalculation on a VID basis.

[0077] This disclosure extracts trace tokens containing identifiers for the principal and elimination columns from the intermediate process of boundary matrix reduction in persistent homology analysis. By referencing the VID assigned to a single constituent vertex corresponding to the trace token via mapping information, the inference state or node state that contributed to the phase change is identified on a VID basis. This enables a configuration in which the internal information of boundary matrix reduction is repurposed for governance control of the distributed inference environment.

[0078] In response to changes in external input or updates to the inference state, the principal column in the boundary matrix reduction process is identified as a column identifier, and VIDs associated with the individual components (including vertices) represented by the column identifier are extracted based on the association information to calculate the contribution, and the set of influential VIDs is determined based on the contribution. Here, the trace token extracted from the reduction process may be configured as identification information including the principal column ID, the deleted column ID, and a computation generation tag (including the generation number) corresponding to the incremental update, and the column correspondence subject to contribution calculation may be uniquely tracked based on the trace token. Furthermore, the rules for determining the set of influential VIDs (threshold and upper k+ confidence adjustment) may be fixed by dependency.

[0079] Here, in the boundary matrix reduction process, the columns that contributed to the determination of birth, death, or change, and the columns that were referenced, added, erased, or replaced during the reduction process of said columns, represent the group of simplices that directly or indirectly contributed to the change in topological features. Each column corresponds to at least one simplice, and since the simplice is constructed based on an inference state, node state, partial representation, or a relationship thereto, at least one VID is associated with the origin or constituent vertex of the simplice. In this disclosure, the column identifiers that contributed to the determination of birth, death, or change, and the column identifiers or origin sets that were involved during the reduction process of said columns are held as trace tokens, and by referring to the VID correspondence information assigned to the origin or constituent vertex of the simplice corresponding to the trace token, the inference states or node states that contributed to the topological change can be reconstructed on a VID basis. This allows this disclosure to identify the influential VID set based on the extent to which it substantially contributed to the change in topological structure, rather than merely a value difference or local numerical fluctuation.

[0080] Therefore, according to this disclosure, components that do not substantially affect the maintenance of the topological structure despite having a large absolute difference can be excluded from the local recalculation target, while components that contribute to changes in topological features through changes in neighborhood relations, connection relations, or cluster structure can be included in the influencing VID set even when the absolute difference is relatively small. Thus, it is possible to achieve rational localization in response to changes in the topological structure while suppressing excesses or deficiencies in the local recalculation target.

[0081] Furthermore, to accommodate cases where there is a many-to-many correspondence between constituent elements (including vertices) and VIDs, the vertices used for contribution aggregation or allocation and the VID assignment rules (equal distribution, aggregation, or reliability weights) can be fixed by dependency.

[0082] Instruct the computation nodes corresponding to the affected VID set to perform local recalculations.

[0083] The surplus resources equivalent to the recalculations avoided by local recalculations are calculated as SavedCost, converted to the number of search branches or beam width using the resource / logic conversion coefficient η, and reallocated only to incomplete branches of the search frontier. The reallocation is then confirmed by an incompleteness confirmation condition or an acceptance response. Here, η can be estimated or calibrated as a conversion coefficient based on the effective cost for each type of operation, or the average additional cost per step of the search (1 branch expansion or 1 beam width increment). The control input for resource conversion reallocation may be the temporal gradient of topological entropy or the magnitude of the temporal stagnation. Furthermore, SavedCost may be reallocated to the dynamic expansion of the prediction horizon, which defines the number of time steps or spatial range to be predicted.

[0084] Here, the amount of physical computational resources calculated as SavedCost has different resource characteristics depending on the type of operation being performed. For example, the inference process in a Transformer-based inference model involves large-scale floating-point operations and memory access, including self-attention mechanisms, linear transformations within layers, and referencing or updating key-value pairs, and can consume large computational resources depending on the model size, input length, growth, and implementation conditions. On the other hand, the boundary matrix reduction process in the persistent homology analysis in this disclosure is configured as a column operation on the Z2 coefficients in at least one embodiment, and at least a part of the addition process can be implemented as an exclusive OR operation or other bitwise operations. Therefore, this disclosure can quantify the resources saved by avoiding recalculation as SavedCost, taking into account the different computational resource characteristics of the inference process and the boundary matrix reduction process, and convert and allocate said SavedCost to search resources. However, since the execution time or resource consumption of the boundary matrix reduction process may depend on the density, column representation, principal structure, and memory access efficiency of the boundary matrix, the calculation of SavedCost is not limited to a single operation, but may be based on at least one of the number of bit operations, memory access amount, communication amount, or execution time.

[0085] An upper limit or an estimated value of the error caused by the phase analysis is calculated as error boundary information, reliability information based on this error boundary information is maintained, and if the reliability is below a threshold, the identification of the influencing VID set is stopped and the process is extended to wide-area recalculation, and the analysis resolution is switched in stages according to the governance signal.

[0086] Hysteresis is applied to the execution and cessation of switching between local and global processes or to the execution and cessation of specific affected VID sets. The process is expanded according to data health, and resource conversion allocation is stopped or reduced if health deteriorates, while the system is evacuated to global recalculation.

[0087] If missing information exists, missing information imputation estimation is performed, and the uncertainty resulting from the imputation is reflected in the error boundary information or reliability information.

[0088] The safety conditions and decision parameters are corrected using external uncertainty indicators or behavioral indicators.

[0089] The system can be moved to an external state storage device 30 (Figure 1), a restart anchor can be generated, and based on the notification information, local / wide-area / cancellation can be selected. ReleasedCost can be added to or combined with SavedCost and treated as surplus capacity for resource recycling.

[0090] One configuration of this device includes assigning a version number, schema identifier, and idempotency key to the allocation confirmation message from the resource conversion and allocation unit 15 (Figure 1). This physically prevents the same resource redistribution process from being applied redundantly, even if retransmission or duplicate reception occurs on the network, thereby ensuring consistency of the resource state in the distributed inference environment.

[0091] The difference determination unit 13 (Figure 1) manages the correspondence between the column identifiers (IDs of the source column and the deleted column), the calculation generation number, and the origin set (Provenance), which is the simple set that forms the basis of the topology change, extracted during the internal calculation process of the topology analysis. This makes it possible to physically track which calculation step is responsible for a change in a particular inference state across multiple distributed computing nodes, and audits the inference consistency across the entire system.

[0092] <Effects of this disclosure> By calculating the contribution based on column identifiers in boundary matrix reduction processing, the influencing VID set can be identified, enabling local recalculation limited to the necessary range, thus improving computational efficiency and responsiveness.

[0093] By fixing the rules for determining the influencing VID set (at least one of the threshold and the top k + reliability adjustment) and the rules for assigning vertices to VIDs (equal distribution, aggregation, or reliability weight) through dependency, it is possible to clarify the differences from previous examples while suppressing inconsistencies in interpretation during implementation. In addition, by configuring the system to reduce the threshold and increase k when reliability deteriorates, the adjustment policy toward the maintenance side (leakage suppression) is clarified.

[0094] (Supplementary information on effects and benefits) This disclosure describes a control system that identifies a set of influential VIDs from trace tokens extracted based on boundary matrix reduction, quantifies the computational resources saved by local recalculation as SavedCost, and converts this SavedCost into search resources using a coefficient η. This enables resource conversion control according to the phase analysis results. The surplus SavedCost, equivalent to the avoided recalculations, can be converted into the number of search branches or beam width and reallocated to incomplete branches, thereby improving search efficiency within the same inference task.

[0095] Furthermore, according to this disclosure, in the boundary matrix reduction process, the group of simplices that contributed to the birth, disappearance, or change of topological features is retained as trace tokens, and by extracting the VID corresponding to the generator or constituent vertex of the group of simplices, the target of local recalculation can be limited to a topologically meaningful range. Therefore, compared to the case where localization is performed based solely on value differences, it is possible to suppress the excess or deficiency of the recalculation target while achieving rational recalculation control in response to changes in the topological structure.

[0096] Furthermore, according to this disclosure, boundary VIDs located outside the set of influential VIDs can be fixed as boundary conditions, and only the interior of the set of influential VIDs can be recalculated. After local recalculation, consistency metrics, residuals, likelihood differences, inference signature differences, or topological feature differences can be evaluated, and a switch to global recalculation can be made as needed. This makes it possible to achieve recalculation that maintains the convergence and consistency of the inference state while suppressing the expansion of inconsistencies due to localization.

[0097] Therefore, this disclosure can reduce computing resources by suppressing unnecessary wide-area recalculations, and improve inference efficiency and search performance in a distributed inference environment by reallocating the resulting freed-up resources to the number of search branches, beam width, search depth, or additional iterations.

[0098] By implementing safe-side transitions based on error boundary information and reliability information, stepwise switching of analysis resolution, and stabilization through hysteresis, operation becomes possible while suppressing misallocation and instability.

[0099] Missing data imputation estimation ensures the continuity of analysis updates and allows the uncertainty caused by imputation to be reflected as reliability.

[0100] By providing a fallback package for evacuation, notification, and restart in the event of failure of local nuclear recalculations, as well as a ReleasedCost cycle, continuous operation can be improved.

[0101] As one aspect of communication control for this device, packet retransmissions and duplicate receipts occurring on the communication path are detected using common header information including version number and schema identifier, as well as an idempotent key unique to each processing request. This physically eliminates the duplicate execution of SavedCost transfer allocation instructions, etc., on the same node, thereby preventing over-allocation of computing resources and inference chattering.

[0102] Column IDs, origin tracking, and generation management clarify the implementation of column identifier tracking, improving auditability and implementability. In addition, a framework is provided to safely utilize the surplus power generated by localization within the same inference task, enabling both improved local recalculation efficiency and enhanced search capabilities as part of a set of governance controls.

[0103] <Definition of Terms> (Lower and upper limits of additional branching) ΔB min ΔB is the lower limit of the number of additional branches ΔB that can be assigned in the search resource reallocation based on SavedCost, and is set to, for example, 0 or 1. max This is the upper limit of the additional branch number ΔB, for example, the safety upper limit (B) of the beam width or number of branches. max ) is set as. The control unit is,

number

[0104] W sc This is a weighting coefficient for reflecting SavedCost in the integrated score, and can be adaptively updated, for example, according to the time window mean or variance of SavedCost.

[0105] The Betti number is an index representing the rank of the homology group corresponding to the persistence diagram or simplicial complex, and includes, for example, β0 (number of connected components) and β1 (number of voids).

[0106] Message Authentication Code (MAC / HMAC) A MAC is an authentication code for verifying the integrity and authenticity of communication messages, and includes methods such as HMAC.

[0107] (Local Recalculation Instruction Packet and Common Header) The local recalculation instruction packet includes at least (i) Version number proto ver (ii) Schema identifier schema ID (iii) Idempotent key IDem key (iv) Issuer identifier issuer ID (v) Timestamp ts (vi) Expiration time ttl (vii) Provenance hash provenance hash (viii) Target VID, column VID list (ix) Additional divergence number ΔB, (x) Message authentication code hmac may include.

[0108] The receiving side can apply idempotently based on IDem and garbage collect processed records due to the expiration of ttl. As an example of the common header, key can be used. As an example of the payload,

Number

Number

[0109] This means that even if local recalculation instructions are resent or duplicated, IDem key This process is idempotent and can be traced through audit logs.

[0110] (Inference state) An intermediate or local state in the inference process that can be described by at least a feature vector, constraint satisfaction information, inference dependency structure, and virtual node identifier VID.

[0111] (High-dimensional state representation) refers to a data representation that integrates the inference state as multiple representation elements (at least feature vectors, constraint satisfaction information, inference dependency structure, and virtual node identifier VID). Furthermore, the high-dimensional state representation may include spatiotemporal coordinates that map the temporal delay associated with the inference process as a physical cost. Here, temporal delay refers to a state in which the temporal delay (or its cost) accumulates relatively large in a local region of a particular VID or inference dependency structure. Note that in this patent, the temporal delay (interaction latency) is not treated as a control trigger for the UI and UX, but is used only as a dimension of the spatiotemporal topology for analyzing the resource allocation shape.

[0112] (Feature vector) A vector representation based on the internal representation of the inference state, attention weights, embeddings, or the summaries / mapping results thereof.

[0113] (Constraint satisfaction information) This refers to information indicating whether a constraint is met or not, whether it is undetermined, contradictory, or to what extent it is satisfied.

[0114] (Inference dependency structure) A structure that represents the dependency or influence relationship between inference states, and may include graphs, adjacency matrices, adjacency lists, reference relationships, or their embedded representations.

[0115] (Virtual Node Identifier VID) refers to an identifier that logically distinguishes inference states in a distributed inference environment and can be associated with compute nodes (physical nodes). The VID may be defined on a logical address space (LogicalAddressSpace) abstracted from physical computing resources. The correspondence between VID and compute nodes is not limited to one-to-one and may change over time. The VID may be implemented as an inference state ID, a representative point ID of the state representation, or an identifier equivalent thereto, and the name or format of the identifier is not limited. Furthermore, the inference state may be represented as a vector representation of intermediate states generated during the inference process (including embedded vectors, hidden state vectors, key vectors or value vectors in the attention mechanism, or cache data holding these, or quantized representations thereof), and the VID may be associated with such vector representation, or the storage location or reference key of such vector representation. Specifically, the phase analysis unit may decide to recalculate the value vector or suppress memory transfer by determining the invariance of the distribution structure (topological structure) of the attention score formed by the set of key vectors.

[0116] (Filtration) This refers to a construction procedure that grows simplices, complexes, etc., stepwise based on changes in distance scales, neighborhood rules, or thresholds for a set of inference states.

[0117] (Reinforcement Definition) The topological entropy in this disclosure is a probability distribution P obtained by normalizing the duration of each topological feature included in the duration diagram. i This includes the Shannon entropy, which is calculated as follows (1).

[0118]

number

[0119] Here, the probability distribution Pi for each duration length li is defined by equation (2).

[0120]

number

[0121] This disclosure is not limited to the definition provided, but includes at least the representative formula. Furthermore, in calculating the phase entropy, normalization may be performed after removing noise components below a specific duration threshold.

[0122] (Persistence Homology Analysis / Topological Features) This refers to an analysis and its results that represent topological properties such as connectivity and cyclicity at each stage of filtration as birth / extinction pairs, duration length, persistence diagram, landscape, persistence image, etc. Topological entropy may be defined as a measure calculated by normalizing the aforementioned topological features (e.g., a set of points in a persistence diagram, or a set of duration lengths) to a probability distribution, and calculating it as Shannon entropy, Renyi entropy, or a statistical complexity equivalent thereto. Topological distortion may be defined as a quantity that represents the local bias or deformation of the distribution of topological features due to temporal stagnation in the aforementioned spatiotemporal coordinates.

[0123] (Reinforcement Definition) A trace token is a record unit that encodes columns that contributed to the determination of birth, extinction, or change in the boundary matrix reduction process (column reduction), and columns that were referenced or erased in the reduction process of said columns, in a form that can be used for subsequent contribution aggregation. A trace token must include at least (i) a pivot column identifier (pivotcolumnID) and (ii) a group of identifiers for erased columns that were added for the erasure process on the column to be reduced in the reduction process, and may optionally include (iii) a set of origin (provenance) of the pivot column and erased columns, (iv) a single identifier, (v) a column ID (single ID) and a generation number, (vi) coefficient information, and (vii) a generation tag. The difference determination unit 13 maps the column set identified by the column identifier or origin set contained in the trace token to a single ID, obtains the vertex set corresponding to the single ID, and further maps it to a VID set using an assignment table (VertexVIDMap) that references the virtual node identifier VID assigned to the vertex (or the inference state corresponding to the vertex). The difference determination unit 13 applies weights to the VID set based on duration, change in topological entropy, or change in topological feature quantity to accumulate the contribution of each VID, and can determine the influential VID set as the VIDs whose contribution is above a threshold, the top k VIDs in terms of contribution, or the logical OR of both. When reliability is reduced, adjustments may be made in the direction of reducing the threshold and increasing k in order to suppress leaks.

[0124] (Boundary matrix reduction) This refers to the process of determining birth and extinction pairs by reducing the matrix representation corresponding to the boundary operator of a simplicial complex.

[0125] (Column Identifier) ​​A column identifier refers to a column ID (single ID) assigned to each column (single) of the boundary matrix, or an identifier that can reference such a column ID (single ID). A column identifier includes the pivot column involved in determining birth / destruction pairs or changes, and may optionally include columns added for elimination during the reduction process of the boundary matrix (eliminated columns). Furthermore, a column identifier may include a set of column IDs (single IDs) included in the provenance set, or a set of columns identified by a record that can reference such a set.

[0126] The column ID (single ID) is an identifier assigned to each column (single) of the boundary matrix at the time of generation, and the generation number is a number that identifies the generation of column additions, replacements, and deletions in incremental updates, etc. Column IDs are generally not reused and may be made traceable in combination with generation numbers.

[0127] (Provenance set) Provenance refers to the origin information that indicates which combination of columns a given column originates from, following a reduction operation (such as column addition) of the boundary matrix. The provenance set may be represented as a set of column IDs, a set of pairs of column IDs and coefficients, or a pointer to a referenced record.

[0128] (Contribution) This refers to the aggregated value of each VID's contribution to the phase change.

[0129] (Influence VID set) This refers to the set of VIDs selected as targets for local recalculation based on changes in topological features. Here, the changes in topological features may be evaluated based on column identifiers, elimination column identifiers, principal column identifiers, or contributions of the simplicial group corresponding to the origin set identified in the boundary matrix reduction process. That is, the influence VID set may be determined as a set containing VIDs corresponding to the generators or constituent vertices of the simplicial group that contributed to the birth, extinction, or change of the topological features.

[0130] Local recalculation refers to recalculating on inference states or node states belonging to the influence VID set. During local recalculation, states corresponding to boundary VIDs located outside the influence VID set may be fixed as boundary conditions. Global recalculation refers to recalculating on multiple nodes, multiple inference states, or the entire system beyond the scope of the influence VID set. The recalculation control unit 14 (Figure 1) may switch from local recalculation to global recalculation as a safer transition if the consistency index, residuals, likelihood difference, inference signature difference, or topological feature difference after local recalculation does not satisfy predetermined conditions.

[0131] (Error boundary information / reliability information) This refers to information representing the upper bound of errors resulting from summarization, approximation, distributed updating, analytical resolution control, or missing value imputation estimation, or the reliability based on such information.

[0132] (Reliability Score / Alternative Reliability Score) The reliability score is a score that represents quality, success rate, resolution, etc., obtained in conjunction with an external device or stored data. The alternative reliability score is a score calculated internally by the device when the reliability score cannot be obtained.

[0133] (Data health) This refers to the state of validity of data used in local analysis and resource allocation, determined based on reliability information, error boundary information, reliability score, or alternative reliability score, etc.

[0134] (Safe transition) This refers to a conservative control selection, such as discontinuing the identification of the VID set and switching to a wide-area recalculation, based on error boundary information, external uncertainty indicators, behavior-related indicators, reliability scores, or data health.

[0135] (Search Frontier / Incomplete Branches) In a search within the same inference task, the set or boundary of incomplete candidate branches is called the search frontier, and the branches that make up that frontier are called incomplete branches.

[0136] (Inference Inconsistency) Inference inconsistency refers to a state in which the inference result after a change in external input or an update of the inference state does not satisfy at least one of the consistency conditions, reference consistency conditions, or constraint consistency conditions with respect to the set of inference states before the update, or a tendency to reach such a state. Inference inconsistency can be observed as at least one of the following: the occurrence of a constraint violation, instability of the result for the same input, or the occurrence of a contradiction after a reference update.

[0137] (Constraint Violation) A constraint violation refers to a state in which a constraint that should be satisfied by the constraint satisfaction information is not satisfied, a contradiction is detected, or the degree of satisfaction falls below a predetermined threshold.

[0138] (SavedCost / ReleasedCost) SavedCost refers to the recalculation equivalent to the cost of recalculations avoided by local recalculations, and ReleasedCost refers to resources released by external backup, etc. These can be defined by at least one of the following: execution time, FLOPs, memory bandwidth, communication bandwidth, etc.

[0139] (Static data) This refers to stored data that can be referenced across sessions for the purpose of backing up and resuming the inference state, and is distinct from volatile intermediate state or temporary data generated during inference execution. Static data may include, but is not limited to, at least a portion of the intermediate inference results associated with the VID, version numbers, schema identifiers, compatibility information, idempotent keys or identifiers, and confidence scores associated with those intermediate inference results.

[0140] (Reinforcement Definition) The unit of the resource / logical conversion coefficient η is defined as "number of search branches / computation cost units". The additional number of branches ΔB is calculated by the following equation (3).

[0141]

number

[0142] If necessary, upper limit B max Apply saturation processing using the specified method.

[0143] (Resource / Logical Transformation Coefficient η) refers to a coefficient for converting the available resources, expressed as SavedCost or ReleasedCost, etc., into the number of search branches or beam width. η can be estimated or calibrated based on the effective computation cost for each type of operation, or the average additional cost per step of the search. Here, in this disclosure, a decreasing function means a function whose output value tends to decrease as the input value increases, and includes, but is not limited to, a monotonically non-increasing function (a function whose output value is always non-increasing as the input value increases). A decreasing function may include a function that has a global decreasing trend but allows local non-monotonicity (for example, a function that applies a moving average, exponential moving average, or low-pass filter to the input value and then applies a monotonically non-increasing transformation), a piecewise decreasing function that decreases in each interval (for example, a step-down function that gradually reduces the value), or a function in which the correlation between the input value and the output value is negative in a given time window.

[0144] (Governance signals) These are signals referenced to control the switching of analysis resolution, the feasibility of identifying impact sets, and the necessity of restraining safe transitions or resource diversions. They may include predictive indicators, load and delay indicators, etc., calculated by external devices.

[0145] (Missing information imputation estimation) This refers to the process of estimating missing information in an inference state group or inference dependency structure based on dependency regression, neighborhood consistency, constraint consistency, etc., and using it for incremental updates in topology analysis.

[0146] (Common Header Information) This refers to header information that may include message type, version number, time information, idempotent key, correlation identifier, etc., in order to ensure interoperability with external devices. The item names, order, expression format, and whether or not they are included in the common header information are not limited.

[0147] (Idempotent treatment) This refers to a treatment that, even if identical or substantially identical messages are resent or received multiple times, does not apply the processing initiated or confirmed by such messages multiple times.

[0148] (UID / TID / UTID, Phase Transition Profile, Inference Signature) UID is an identifier that identifies the unit being managed in the inference process, and refers to an identifier that is associated with, for example, the configuration of the inference model, the type of inference task, and the inference environment (computation node group, data source, version number, etc.). TID refers to an identifier that identifies the unit of an inference task or inference execution. UTID is an identifier that identifies the unit of inference execution when inference is performed in association with a user's input generation process or interactive operation, and may be associated with at least one of the UID and TID. A phase transition profile refers to a profile that stores at least one of topological features, contributions, influence VID sets, error boundary information or reliability information, SavedCost and ReleasedCost, or parameters or derived indices based thereon, in association with the UID, TID, or UTID over time. An inference signature is an index that represents a set of logical state transitions in an inference process, identified or updated based on a phase transition profile, and may include, for example, a persistence diagram, a barcode, a landscape, a path signature, or at least one of these summaries (e.g., persistence length distribution, distance statistics). Inference signatures are limited to uses referenced for the governance of a distributed inference environment, i.e., determining the scope of local recalculation, switching between local and global, switching analysis resolution, resource allocation of SavedCost and ReleasedCost, or auditing and consistency verification, and are not limited to uses for personal attribute estimation, psychological or medical diagnosis, or personal identification.

[0149] (Distance scale D / Neighborhood rule R / Threshold sequence Θ / Signature spatial distance d) sig The continuity threshold ε) and distance scale D refer to a measure that defines the distance between higher-dimensional state representations. The neighborhood rule R refers to a rule that defines neighborhood relations (e.g., k-neighborhoods, ε-neighborhoods, etc.) based on D. The threshold sequence Θ refers to a sequence or change procedure of thresholds (distance thresholds, neighborhood thresholds, or thresholds for join rules, etc.) that define the stages of filtration. The names, forms, and inclusions of D, R, and Θ are not limited. Furthermore, the signature spatial distance d sig This refers to a measure that defines the distance between inference signatures. sigThis can be defined, for example, as the bottleneck distance, Wasserstein distance, landscape distance, or a statistic summarizing such distance between persistence diagrams. The continuity threshold ε is the allowable distance for considering the inference signature to have changed continuously along a time series, for example, the inference signature S corresponding to consecutive time windows k, k+1. k S k+1 This can be defined as a threshold for treating a transformation as continuous if it satisfies the following equation (4).

[0150]

number

[0151] <One Embodiment of the Present Disclosure> Hereinafter, embodiments for implementing this disclosure will be described with reference to the attached drawings. The embodiments described below are all preferred examples of this disclosure. Therefore, the numerical values, shapes, materials, components, and their arrangement and connection configurations shown in the following embodiments are examples and are not intended to limit this disclosure. Accordingly, components in the following embodiments that are not described in the independent claims representing the highest-level concepts of this disclosure will be described as arbitrary components. Furthermore, in each figure, substantially identical components are denoted by the same reference numerals, and redundant explanations are omitted or simplified.

[0152] (Device configuration) First, with reference to Figure 1, the configuration of an information processing device 10 according to one embodiment of this disclosure will be described. The information processing device 10 includes a state acquisition unit 11, a phase analysis unit 12, a difference determination unit 13, and a recalculation control unit 14. Furthermore, it may include a resource conversion and allocation unit 15 and a surplus capacity calculation unit 19. Optionally, it may also include an index receiving unit, a transfer unit 16, a notification receiving unit 17, a profile table 18, and an external state storage device 30. The computing node may include non-volatile memory (high-bandwidth flash memory, etc.) located in close proximity to the arithmetic unit (GPU, etc.). The uncertainty index or behavior-related index received by the index receiving unit may correspond to an index calculated by an external device that evaluates behavioral uncertainty in the dialogue input. In this case, an increase in behavioral uncertainty in the external device is reflected in the correction of the safe-side transition conditions or determination parameters of this device, enabling adaptive control of the distributed inference environment in response to fluctuations in the dialogue input.

[0153] The information processing device 10 may be implemented in a single device or distributed across multiple devices.

[0154] (Acquiring the inference state (high-dimensional state representation)) The state acquisition unit 11 acquires the inference state during the inference process as a high-dimensional state representation.

[0155] The high-dimensional state representation includes at least a feature vector, constraint satisfaction information, an inference dependency structure, and a virtual node identifier (VID). The correspondence between VID and computation node is not limited to one-to-one and may vary over time.

[0156] At least one of the mapping pieces used to associate individual components (including vertices) with inference states (or their representative points) and VIDs may be generated during the configuration and / or update of the filtration and may be held as a table, a reference pointer, or a referenceable record.

[0157] The aforementioned mapping information is maintained with generation tags that match the column ID (single ID) and generation number. When vertex IDs or VIDs are added, replaced, deleted, or reassigned due to incremental updates, the mapping (column identifier, single ID, vertex, and VID) is resolved only for references that share the same generation tag, thereby avoiding or suppressing mappings based on the mixing of different generation tags.

[0158] In incremental updates, (1) New IDs may be assigned to any additional vertex IDs or VIDs, and the reuse of existing IDs may be prohibited or avoided. (2) When a vertex ID or VID is to be replaced, the old ID is retained as a tombstone and replaced with the new ID. (3) The vertex ID or VID that is deleted will be retained as expired and will not be subject to reference resolution. (4) The column ID (single ID) and generation number are updated to match the generation tag of the update, and the mapping is resolved only for references where the generation tag matches. (5) The origin set (provenance) and the results of the correspondence resolution can be recorded with generation tags to avoid or prevent the mixing of different generations. (6) If a request is received that references an expired (tombstone), the identification of the affected VID set may be stopped as a safe transition, and the scope of recalculation may be expanded to a wider area.

[0159] Filtration and persistent homology analysis (run / incremental update, resolution switching, error boundaries) The phase analysis unit 12 constructs a filtration for the set of high-dimensional state representations and calculates topological features by performing or incrementally updating persistent homology analysis.

[0160] Each column (individual) in the boundary matrix may be assigned and retained a column ID (individual ID) at the time of generation. If columns are added, replaced, or deleted through incremental updates, the reuse of column IDs may be prohibited, and a generation number may be assigned to ensure consistency in column correspondence. The identification of column identifiers and the aggregation of contributions may be traceable using column IDs and generation numbers.

[0161] When operations such as column addition are performed during the reduction process, the phase analysis unit 12 or the difference determination unit 13 may propagate and record the origin set (provenance). The origin set is, for example, (a) Set of column IDs, (b) A set of pairs of column IDs and coefficients, or (c) Reference to a record that can refer to the set It can be expressed as follows.

[0162] The difference determination unit 13 may use a set of column IDs (single IDs) included in the origin set (provenance), or a set of columns identified by a record that can reference said set, as a column identifier, without being limited to the original column or the deleted column. This makes it possible to track the set of columns that contributed to the determination of the birth / destruction pair, even if there are inconsistencies in column identification due to differences in the implementation of the reduction operation (for example, differences in the order of operations, the order of distributed updates, or the internal representation).

[0163] The distance scale D and the neighboring rule R are not limited, but may be, for example, a weighted and combined distance obtained by combining the distance between feature vectors and the difference in constraint satisfaction information. Filtration may be structured as, for example, a stage in which the distance threshold is monotonically changed as a threshold sequence Θ, a stage in which the neighboring rule R (k-neighbors, ε-neighbors, etc.) is changed stepwise, or a stage in which the combined rule based on the inference-dependent structure is relaxed stepwise.

[0164] When errors are introduced through summarization, approximation, distributed updating, or reduction of analysis resolution, the phase analysis unit 12 may calculate an upper limit of the error or an estimated value of the upper limit as error boundary information and store reliability information based on the error boundary information in association with the VID.

[0165] The summarization rate, approximation rate, distributed update delay, and missing value used in calculating the error boundary information are all numerical indicators that can cause a decrease in accuracy during the execution of phase analysis or incremental updates, and are used individually or in combination to estimate the upper bound of the error. Since the available indicators may differ depending on the configuration and operating conditions of the distributed inference environment, adaptability to diverse operating environments is ensured by configuring the error boundary information to be calculated based on at least one of the available indicators.

[0166] The phase analysis unit 12 may switch the analysis resolution in stages according to the governance signal, normally updating the phase features based on low-resolution summary features, and switching to high-resolution estimation and updating when predetermined conditions are met.

[0167] (Identification of the set of influencing VIDs (sequential correspondence between column identifiers, individual elements, vertices, VIDs, and contributions)) The difference determination unit 13, triggered by a change in external input or an update in the inference state, extracts trace tokens from the reduction process of the boundary matrix in persistent homology analysis, and identifies the influencing VID set by a mapping chain that sequentially links column identifiers, symmetrics, vertices, and VIDs based on the trace tokens.

[0168] The difference determination unit 13 identifies the column that became the pivot when the birth / destruction pair was determined by the reduction process as the pivot column, and stores the column ID (single ID) and generation number of the pivot column as column identifiers.

[0169] In the process of identifying the scope of influence by the difference determination unit 13, in addition to the principal column in the reduction process of the boundary matrix, the identifiers of columns added for the elimination process on the columns to be reduced during the reduction process (eliminated columns) may be identified as column identifiers and used to calculate the contribution to the topology change. Here, the difference determination unit 13 is characterized by performing a pruning process that excludes column identifiers whose duration length of the topological feature is less than a predetermined noise threshold from the target of trace token retention, and selectively tracking the origin set (Provenance) that contributes to macroscopic topology changes.

[0170] The difference determination unit 13 can treat a trace token including a column identifier, an elimination column identifier, a pivot column identifier, or a derived set extracted from the reduction process of the boundary matrix as a record of a simple group that contributed to the change in the topological feature amount. Each column of the boundary matrix corresponds to at least one simplex, and the simplex is constructed based on an inference state, an intermediate inference state, a node state, or a dependency or neighborhood relationship between them. Therefore, one or more VIDs are associated with each simplex via its generator or constituent vertex.

[0171] In the reduction process of the boundary matrix, when a certain column functions as a pivot column that determines the birth or disappearance, the simplex corresponding to the column is a simplex that participated in the generation, disappearance, or lifespan of at least one topological feature amount. Further, the columns involved in reference, addition, elimination, or replacement during the reduction of the pivot column correspond to the simplexes that contributed to the local structure for determining the change in the topological feature amount. For this reason, the difference determination unit 13 can extract a local region sensitive to the change in the topological structure by holding, as trace tokens, not only the pivot column itself but also the column group involved in the reduction of the pivot column.

[0172] Furthermore, the difference determination unit 13 can restore the group of inference states that contributed to the change in the topological structure as a VID set by referring to the generator, constituent vertex, or state identifier referred to at the time of simplex generation of the simplex corresponding to the trace token. In this way, the present disclosure can accurately extract the inference state or node state that contributed to the change in the topological feature amount by paying attention not only to the result of the boundary matrix reduction itself but also to the contribution history of which column participated in the transformation of which column in the reduction process.

[0173] Here, in a configuration that determines a locally recalculated target based only on a simple value difference, there may be a case where components with a large value difference but that do not contribute to the phase structure are overly selected. On the other hand, there may be a case where components that contribute to the change in the phase feature amount through the change in the connection relationship or the neighborhood relationship are overlooked despite having a small value difference. In contrast, in the present disclosure, in order to extract an influence VID set through a monomer group that substantially contributes to the fluctuation in the birth, disappearance, or lifespan of the phase feature amount, it is possible to limit the recalculation range to a range that is phase-wise meaningful while suppressing the excess or deficiency of the locally recalculated target.

[0174] Further, the difference determination unit 13 may calculate a contribution score for each VID based on the contribution degree per monomer, the lifespan difference, the persistence diagram distance, or the number of contributions in the simplification process. In this case, the influence VID set may be determined as a set including VIDs whose contribution scores are equal to or greater than a predetermined threshold, or may be determined as a set including, in a lump, a group of VIDs involved in the same birth / extinction pair or the same change in the phase feature amount. Thereby, the discrete observation results of the simplification trace can be stably mapped to a VID set suitable for local recalculation control.

[0175] The difference determination unit 13 identifies a monomer corresponding to the column identifier and extracts the constituent elements (including vertices) of the monomer.

[0176] The difference determination unit 13 extracts the VID associated with the extracted vertex (or representative point) based on the association information.

[0177] The difference determination unit 13 performs weighting based on the persistence length and aggregates the contribution degrees for each VID.

[0178] When there is a many-to-many relationship between the vertex (constituent element) and the VID, the difference determination unit 13 aggregates or assigns the contribution degrees according to the vertex ⇔ VID assignment rule.

[0179] The difference determination unit 13 determines the influence VID set based on the contribution degree.

[0180] The difference determination unit 13 may adjust at least one of the predetermined threshold and k according to the reliability information, reliability score, surrogate reliability score, or data health.

[0181] The predetermined reliability information, reliability score, alternative reliability score, and data health all share a common technical function as indicators representing the reliability of the phase analysis or inference state. The adjustment of the threshold or k in the difference determination unit 13 is performed according to a control policy that prioritizes leak suppression as the value of the reliability indicator decreases and prioritizes efficiency as it increases. Depending on the implementation availability, connection status with external devices, or operating conditions, it may be selected which indicator to use, or to use multiple indicators in an integrated manner. For example, if a reliability score can be received from an external device, that reliability score may be used preferentially; if it cannot be received, it may be switched to an alternative reliability score; in either case, it may be used in conjunction with the reliability information calculated internally by the device to make an integrated determination as data health.

[0182] The difference determination unit 13 may, in principle, include all VID sets associated with the same birth / destruction pair (or simplex set) in the influencing VID set.

[0183] (Instruction and execution of local recalculation (sequential correspondence between affected VIDs and compute nodes)) The recalculation control unit 14 instructs the computing nodes corresponding to the influencing VID set to perform local recalculations.

[0184] The correspondence between the set of influential VIDs and the compute nodes may be determined based on the VID-to-compute node correspondence table held by the state acquisition unit 11, or metadata obtained from the distributed inference environment.

[0185] Local recalculations can be performed on inference states belonging to the influence VID set.

[0186] When the recalculation control unit 14 performs a local recalculation based on the influence VID set, it may fix the inference state, node state, or intermediate representation corresponding to the boundary VID located outside the influence VID set and directly or indirectly connected to the influence VID set as a boundary condition, and recalculate only the inference states belonging to the inside of the influence VID set as the target for updating. This makes it possible to perform a recalculation while suppressing the propagation of differences within the local region and maintaining consistency with stable inference states outside the local region.

[0187] In other words, if an inference state located outside the set of influencing VIDs is already stable and its contribution to the change in topological features is below a predetermined threshold, the recalculation results within the local region can be consistent with the stable external state even if that inference state is excluded from the recalculation. For this reason, the local recalculation of this disclosure does not simply reduce the amount of data to be processed, but functions as a conditional recalculation with a fixed outer boundary, focusing on the range that contributed to the change in the topological structure.

[0188] The recalculation control unit 14 evaluates at least one of the following criteria for the inference result obtained after local recalculation: topological feature difference, persistence distance, inference signature difference, residual, likelihood difference, consistency index, or convergence criterion. If the evaluation result satisfies predetermined conditions, the local recalculation result may be adopted. This allows the disclosure to determine whether the state after local recalculation falls within an acceptable range relative to the original global state or the expected consistent state.

[0189] On the other hand, if the size of the influencing VID set exceeds a predetermined ratio, if the topological feature difference, persistence distance, inference signature difference, residual, likelihood difference, consistency index, or convergence criterion does not meet the predetermined conditions after local recalculation, or if the residual or difference does not decay even after a predetermined number of iterations, the recalculation control unit 14 may cancel the local recalculation and switch to global recalculation. This allows the present disclosure to safely transition to global recalculation when necessary, while suppressing the fixation of inconsistencies or accumulation of errors caused by the application of local recalculation.

[0190] Therefore, according to this disclosure, local recalculation based on the influence VID set is implemented as a control that combines boundary condition fixing, local consistency evaluation, and fallback to global recalculation, making it easier to perform local recalculations while maintaining the convergence of the inference state. This makes it possible to ensure the consistency and stability of inference results while suppressing the frequency of global recalculations and the consumption of computational resources.

[0191] If local recalculation fails, or if an abnormal state preventing recalculation is detected, the system may optionally transition to external backup. The recalculation control unit 14 may also generate a local control command for at least one of the inference state, intermediate representation, cache area, compute node, or communication path corresponding to the influential VID set, based on the result of identifying the influential VID set. This local control command is not merely a recalculation instruction, but aims to maintain consistency across the entire distributed inference environment and resolve inconsistencies early. Specifically, the recalculation control unit 14 may instruct a "temporary suspension of output reflection," which temporarily suspends the presentation of inference results related to the influential VID set to subsequent processes or externally until local recalculation is completed and the consistency of the inference is confirmed. Furthermore, in order to resolve the inconsistency in the local region, the compute node responsible for processing the affected VID set may be dynamically instructed to "change the processing priority (e.g., raise the priority)" of the task, to "change the allocation of compute resources (allocate additional compute cores or memory)" to cope with a temporary increase in computing load, and to "change the allocation of communication bandwidth (prioritize bandwidth)" to accelerate state synchronization between the relevant nodes. In other words, the local control instruction may include at least one of the following: local recalculation, consistency verification, cache invalidation, state value fixation, temporary suspension of output reflection, change in processing priority, change in allocation of compute resources, change in allocation of communication bandwidth, state saving, issuance of audit identifiers, generation of impact scope records, calculation of hashes for consistency verification, or transition to a global recalculation.

[0192] (Calculation of saved cost and allocation of resources to exploration (number of branches and beam width)) The remaining power calculation unit 19 calculates the remaining power corresponding to the recalculation avoided by the difference determination and local recalculation as SavedCost.

[0193] SavedCost can be defined, for example, as the reduced amount of FLOPs, memory bandwidth usage, communication bandwidth usage, or execution time that has been avoided.

[0194] The resource conversion distribution unit 15 converts SavedCost into an addition term to the additional search branch number or beam width based on the resource-logical conversion coefficient η, and redistributes and applies it limited to the parallel branches belonging to the search frontier identified as unfinished within the same inference task. η can be estimated or calibrated by short-time profiling before the start of inference or during inference execution.

[0195] The resource conversion distribution unit 15 determines the resource conversion distribution on the condition of confirming that the target branch is in an unfinished state or on the condition of the reception response from the target calculation node. If it does not reach a determination, the resource conversion distribution may be invalidated or reassigned to other unfinished branches. If the reassignment does not reach a determination for a predetermined number of consecutive times, all or part of the remaining power may be discarded.

[0196] When it is impossible to determine the unfinishedness, the resource conversion distribution unit 15 does not make a determination for the branch and may retain all or part of the corresponding remaining power for a predetermined period and use it for the next allocation determination.

[0197] In a distributed environment, distribution determination messages, evacuation data, restart packages, etc. may be resent or received repeatedly. In preparation for this, idempotent handling using common header information and idempotent keys may be implemented.

[0198] When the resource conversion and allocation unit 15 determines the reallocation of resources based on saved cost, it sends a resource allocation command to the target computing node that includes the resource conversion type (addition of search branches or increment of beam width), a specific increment value ΔB, a target search frontier identifier, and an idempotent key. The allocation state is determined by triggering the receipt of an acknowledgment (ACK) from the computing node or confirmation that the target branch is in an incomplete state, thereby physically ensuring consistency in asynchronous resource control in a distributed environment. The idempotent key can be generated as a hash including the UTID, target branch identifier, and allocation generation number, etc.

[0199] (Error boundary information, reliability information, safe transition, hysteresis) The phase analysis unit 12 calculates an upper limit or an estimated value of the error introduced due to summarization, approximation, variance update, or reduction of analysis resolution as error boundary information, and stores reliability information based on the error boundary information in association with the VID.

[0200] The difference determination unit 13 may cancel the identification of the affected VID set and expand the recalculation target to a wider area if the reliability information falls below a predetermined reliability threshold.

[0201] The phase analysis unit 12 may switch the analysis resolution in stages according to the governance signal.

[0202] The difference determination unit 13 or the recalculation control unit 14 may suppress chattering by applying a hysteresis characteristic that makes the activation threshold and the recovery threshold different when switching between local mode and wide-area mode, or when executing and stopping the execution of a specific set of influential VIDs. The hysteresis width may be expanded based on data health.

[0203] The resource conversion and allocation unit 15 may temporarily suspend the application of resource conversion and allocation, invalidate undetermined allocations, or reduce the allocation amount if the data integrity does not meet predetermined conditions or if the error boundary information increases by a predetermined amount or more.

[0204] (Integrated reliability score / Alternative reliability score (estimated in case of non-delivery)) The system may receive or obtain a reliability score corresponding to the stored data from the external state storage device 30 or an external device.

[0205] If a reliability score is unavailable, the phase analysis unit 12 may calculate an alternative reliability score based on at least one of the following: error boundary information, missing data rate, degree of approximation, summarization rate, delay amount of distributed update, or estimated error based on short-time profiling.

[0206] The resource conversion and allocation unit 15 may adjust the allocation amount, the number of branches to be allocated, the number of branches to be searched, or the addition term to the beam width, or η to the maintenance side, or suppress the application of the allocation, depending on the reliability score or the alternative reliability score.

[0207] The resource conversion allocation unit 15 may use as a confirmation condition that, before and after the application of resource conversion allocation, the inference result or inference state satisfies a predetermined quality condition based on at least one of constraint satisfaction information, reference consistency conditions, or consistency confirmation conditions, or that the probability of not satisfying said quality condition is below a predetermined value. If the confirmation condition is not met, the application of resource conversion allocation may be suppressed or the allocation amount reduced.

[0208] (Estimation with missing information and reflection of uncertainty) The phase analysis unit 12 may perform missing information imputation estimation if there are missing values ​​in the inference state group or inference-dependent structure.

[0209] Error boundary information or reliability information is updated to reflect the uncertainty resulting from complementary estimations.

[0210] (Adjustment using external uncertainty indicators and behavioral indicators (optional)) The indicator receiving unit may receive uncertainty indicators from an external device.

[0211] The technical meanings of each value included in the aforementioned uncertainty indicators are as follows: Input consistency refers to a value representing the degree to which the external input supplied to the inference model conforms to a predetermined schema, type constraint, or range constraint. Syntactic uncertainty refers to a value based on the degree of parsing ambiguity or the number of alternative parsings that occur in the structural analysis of the input (e.g., generation of a syntax tree, parsing of a token sequence, or determination of conformance to a formal grammar). Semantic uncertainty refers to a value representing the degree of ambiguity in the semantic interpretation of the input, the degree of reference ambiguity, or the degree of context-dependent semantic variation. Constraint satisfaction uncertainty refers to a value representing the degree to which the satisfaction state of the constraints imposed on the inference process is uncertain, or the range of variation in the degree of satisfaction. These values ​​are calculated in the external device during the input preprocessing, analysis, or evaluation process and supplied to the indicator receiving unit of this device.

[0212] The difference determination unit 13 may discontinue the identification of the affected VID set and expand the recalculation target to a wider area if the uncertainty index satisfies predetermined safety conditions.

[0213] The behavior indicator receiving unit may receive behavior-related indicators related to the user's input generation process or dialogue operation.

[0214] The difference determination unit 13 may adjust the threshold, weight, judgment coefficient, or distance threshold of the safety condition in a direction that is on the safety side if the behavior-related indicator satisfies predetermined deterrence conditions. Hysteresis may be applied to set different conditions for deterrence and recovery.

[0215] The transfer unit 16 may transfer the inference state as static data to the external state storage device 30 if local recalculation fails or if an abnormal state that makes recalculation impossible is detected.

[0216] The external state saving device 30 may correspond to a device that saves the state in the event of an inference anomaly and performs selective restart by an update trigger. In this case, the restart anchor generated by the external device based on the intermediate inference results transferred from the transfer unit 16 includes a compressed representation of the confirmed intermediate inference results corresponding to the VIDs belonging to the influence VID set, and the structure identifier may include the computation generation tag and the type identifier of the unsatisfied condition included in the trace token. This allows the results of identifying the influence range based on the phase analysis of this device to be used to limit the search range of the difference calculation in the external device, and the efficiency of local recalculation and the accuracy of inference restart can be improved in conjunction.

[0217] The external state storage device 30 may generate and store a restart anchor, which serves as the starting point for restarting the inference, and a structural identifier that identifies any unsatisfied conditions or scope of influence required for restarting, based on the transferred intermediate inference results, or it may generate a restart package containing the stored restart anchor and structural identifier. The restart package may include a version number, a schema identifier, compatibility information, and an idempotent key or identifier.

[0218] The transfer unit 16 may limit the transfer to intermediate inference results corresponding to VIDs belonging to the influential VID set, or VIDs whose contribution satisfies predetermined conditions.

[0219] The notification receiving unit 17 may receive notification information from the external state storage device 30 indicating whether or not to perform a restart or recalculation, or whether or not to generate or send a restart package. Based on the notification information, the recalculation control unit 14 may select either a local recalculation, a wide-area recalculation, or to cancel the recalculation.

[0220] The resource conversion and allocation unit 15 may define resources released by transfer or evacuation as ReleasedCost, add them to SavedCost, or separately combine them to treat them as surplus, and apply them as an additional term to the number of search branches or beam width.

[0221] (UTID, phase transition profile, inference signature (limited use)) The device may store topological features, contributions, influence VID sets, error boundary information or reliability information, or history or parameters such as SavedCost / ReleasedCost associated with UID, TID, or UTID.

[0222] Phase transition profiles may be accumulated in a time series and stored in a profile table 18, and inference signatures may be identified or updated based on the phase transition profiles. For example, the device may identify an inference signature S for each time window or processing interval. k Generates the S k This may be stored in association with a UID, TID, or UTID.

[0223] The device is a set of reference inference signatures S. ref (For example, representative value, median, or cluster center over a predetermined period) and candidate inference signature S cand (For example, S corresponding to the latest window k ) and may be kept separately.

[0224] The device has a signature spatial distance d sig Based on S cand S ref You may determine whether or not it belongs to an equivalence class or reachable region.

[0225] Here, the reachable region is S ref From S cand The inference signature sequence {S} leading to i} exists, and the column satisfies the continuity condition shown in equation (5) below, and as shown below d sig (S i ,S cand The region defined by at least one of the following conditions—that ) is monotonically decreasing or exhibiting a predetermined decreasing trend—can be defined as shown in equation (6).

[0226]

number

[0227] The device is S cand If it is determined that S belongs to the equivalent class or the reachable region, cand The system may incorporate the existing inference signatures as targets for update, or control the update coefficient, update cycle, or whether or not to perform the update of the inference signatures. Furthermore, the system may adjust the determination rules for the influence VID set by the difference determination unit 13, the switching between local and wide-area analysis by the recalculation control unit 14, the switching of analysis resolution by the phase analysis unit 12, or the allocation determination conditions for SavedCost / ReleasedCost by the resource conversion allocation unit 15, depending on the result of the assignment determination.

[0228] Reference to inference signatures is limited to governance purposes. Specifically, inference signatures may be referenced for the control and auditing of the distributed inference environment, and may be used, for example, to assess the stability of the impact VID set, restrict or expand the scope of local recalculations, determine fallback conditions to global recalculations, stepwise switch of analysis resolution, correct safe transition conditions based on error boundary information and reliability information, determine or invalidate resource conversion allocation for SavedCost / ReleasedCost, or audit duplicate application suppression based on idempotency and common header information.

[0229] Next, an example of the processing flow executed by the information processing device 10 will be explained with reference to Figures 2 to 4.

[0230] The state acquisition unit 11 acquires the inference state group as a high-dimensional state representation and maintains the VID and computation node correspondence. (S2) The phase analysis unit 12 configures filtration and performs or incrementally updates sustained homology analysis. (S3) When an external input is changed or the inference state is updated, the difference determination unit 13 identifies the main source column of the reduction process (including the deletion column as necessary, or a generalized set of columns based on the origin set) as a column identifier. (S4) The difference determination unit 13 extracts VIDs based on the column identifier, individual, vertex, and VID mapping information, and calculates the contribution by weighting based on the duration length. (S5) The difference determination unit 13 determines the set of influential VIDs based on the degree of contribution (adjusted according to the threshold or top k and reliability). (S6) The recalculation control unit 14 instructs the computing node corresponding to the influencing VID set to perform a local recalculation and reflects the results.

[0231] The reserve cost calculation unit 19 calculates the reserve cost equivalent to the recalculation that was avoided by localization as SavedCost. (T2) The resource conversion and allocation unit 15 converts SavedCost into the number of search branches or beam width increments using η. (T3) The resource conversion and allocation unit 15 selects an incomplete branch from the exploration frontier and sends an allocation proposal. (T4) The allocation is determined based on the confirmation condition that the target branch is incomplete, or based on the acknowledgment received from the target node. (T5) If it cannot be determined, it will be invalidated or reallocated, and if it fails for a predetermined number of consecutive times, all or part of the remaining amount will be discarded.

[0232] The phase analysis unit 12 calculates error boundary information and stores reliability information. (U2) If the reliability is below the threshold, the difference determination unit 13 stops identifying the affected VID set and extends to wide-area recalculation. (U3) The analysis resolution is switched in stages according to the governance signal, with low-resolution updates under normal circumstances and switching to high-resolution estimation only when necessary. (U4) Hysteresis is applied to the switching between local and global, and between execution and suspension, to suppress chattering.

[0233] If a missing value is detected, the phase analysis unit 12 performs missing value completion estimation based on dependency regression, neighborhood matching, constraint matching, etc. (V2) The error boundary information is updated based on the interpolation ratio, etc., and corrected in a way that reduces reliability.

[0234] When a local recalculation fails or an anomaly is detected, the transfer unit 16 transfers static data to the external state storage device 30. (W2) The external state storage device 30 generates a restart anchor and a structural identifier, and generates a restart package. (W3) The notification receiving unit 17 receives notification information relating to the restart package, and the recalculation control unit 14 executes at least one control selected from the group consisting of local recalculation, wide-area recalculation, and termination of processing, based on the notification information. (W4) The resource conversion and allocation unit 15 manages the ReleasedCost, which is the resource released due to the suspension or suppression of processing, by adding it to the SavedCost as surplus capacity.

[0235] Each part of this disclosure may be implemented by software (a program to be executed by a computer), hardware, or a combination thereof.

[0236] In one embodiment of this disclosure, the determination of structural isomorphism between multimodals by the topological analysis unit 12 may be performed independently of the identification of the influencing VID set. That is, the configuration in which topological features calculated from latent representations of different modalities are projected onto a common causal graph space, and structural inconsistencies are determined based on the bottleneck distance of persistent barcodes or persistent diagrams in that space, is not limited to distributed inference environments but can also be applied to multimodal inference on a single device. This configuration makes it possible to verify whether the logical consistency of inference is maintained between different modalities such as text, images, and numbers as isomorphism of topological structures, without relying on semantic interpretation, and contributes to the detection of structural hallucination.

[0237] The distance scale D, neighborhood rule R, threshold sequence Θ, type of filtration, and representation format of topological features are not limited. The definitions of SavedCost and ReleasedCost, the method for estimating η, the method for determining incompleteness, and the method for generating idempotent keys are also not limited.

[0238] This disclosure is not limited to distributed inference environments, but can also be applied to parallel inference within a single device or to multi-model cooperative inference.

[0239] <Example 1: Boundary matrix reduction using conversation logs, trace token generation, and identification of the influential VID set> Next, Example 1 will be described with reference to Figures 8 and 9. Although this example illustrates a conversation log, the scope of this disclosure is not limited to this. It is applicable as long as it can be mapped as a point cloud (VID) in a high-dimensional space. This embodiment is a reference example illustrating a specific set of operations on filtration, boundary matrix reduction, trace token generation, and influence VID set, based on a simplified example using a conversation log as the inference state, in order to facilitate understanding of the present disclosure. This embodiment is for conceptual purposes only and may be replaced in practical applications with configurations based on the aforementioned high-dimensional state representation, distance scale D, neighborhood rule R, and inference dependency structure.

[0240] (Definition of data and vertices) In this embodiment, five utterances [v1, v2, v3, v4, v5] were defined as 0-simple (vertices). Furthermore, 1-simple (edges) and 2-simple (triangles) were constructed to reflect the semantic chain relationships between the utterances and the relationships between the three. Filtration values ​​were set progressively in the order of vertices, edges, and triangles according to the strength of the semantic chain (details are shown in the next section). This resulted in a structure where the simplex complex grows progressively as filtration progresses.

[0241] (Details of filtration values) In this embodiment, the filtration value reflecting the strength of semantic sequences was set as follows. The strength of semantic sequences can be defined as an index based on inter-utterance dependencies, reference relationships, or temporal proximity.

[0242] The filtration value of vertices (v1~v5) was set to 0. The filtration value of edges [v1,v3] and [v3,v5], which have strong semantic chains, was set to 1, and the filtration value of other edges [v1,v2], [v2,v3], [v3,v4], and [v4,v5] was set to 2. The filtration value of triangle [v1,v2,v3] was set to 3, and the filtration value of triangle [v3,v4,v5] was set to 4. As a result, the column identifier (σ1~σ) is obtained by arranging the individual elements in order of their filtration values. 13The following was added: ·σ1=v1,σ2=v2,σ3=v3,σ4=v4,σ5=v5(f=0) ·σ6=[v1,v3],σ7=[v3,v5](f=1) ·σ8=[v1,v2],σ9=[v2,v3],σ 10 =[v3,v4],σ 11 =[v4,v5](f=2) ·σ 12 =[v1,v2,v3](f=3) ·σ 13 =[v3,v4,v5](f=4)

[0243] (Construction and reduction of the boundary matrix ∂) The aforementioned single-order sequence σ1~σ 13 Based on this, the boundary matrix ∂ on Z2 was constructed. The non-zero elements of each column j are as follows: • Columns σ1~σ5 (vertices): The boundary is empty (zero column). ·Column σ6(=[v1,v3]): Non-zero in rows σ1, σ3. ·Column σ7(=[v3,v5]): Non-zero in rows σ3, σ5. ·Column σ8(=[v1,v2]): Non-zero in rows σ1, σ2. ·Column σ9(=[v2,v3]): Non-zero in rows σ2, σ3. · Column σ 10 (=[v3,v4]): Non-zero in rows σ3 and σ4. · Column σ 11 (=[v4,v5]): Non-zero in rows σ4 and σ5. · Column σ 12 (=[v1,v2,v3]): Non-zero in rows σ6, σ8, σ9. · Column σ 13 (=[v3,v4,v5]): row σ7,σ 10 ,σ 11 It is not zero.

[0244] The device performs column reduction on the boundary matrix ∂ in ascending order of column numbers on Z2 to calculate the reduced matrix R. During the reduction process, if the lowest 1 (pivot) of a column overlaps with the lowest 1 of a preceding column, the preceding column is added to eliminate the lowest 1. In this process, the column in which the lowest 1 remains is identified as the pivot column, and the column in which all elements become 0 is called the zero column. The column added for the elimination process of the column to be reduced is called the elimination column.

[0245] (Execute column reduction process) Column reduction on Z2 was performed on the boundary matrix in ascending order of column number (Figure 8(b)). The reduction results for each column are as follows. Note that low(j) represents the row number of the non-zero element in the bottom row after reduction of column j. (i) Columns σ1~σ5: Since they are zero columns, no reduction is necessary. Five zero-dimensional features (connected components) are created. (ii) Column σ6: low(σ6)=σ3. No antecedent matrix has the same low row. Confirmed as the principal element column. Record the birth / annihilation pair (σ3,σ6). v3 is born at (f=0) and annihilated by merging with v1 due to the addition of edge [v1,v3] (f=1). Duration length=1. (iii) Column σ7: low(σ7)=σ5. Confirmed as the principal column. Birth / extinction pair (σ5,σ7). v5 is born at (f=0) and extinct at (f=1). Duration = 1. (iv) Column σ8: low(σ8)=σ2. Confirmed as the principal column. Birth / extinction pair (σ2,σ8). v2 is born at (f=0) and extinct at (f=2). Duration = 2. (v) Column σ9: low(σ9)=σ3. Same low row as column σ6. Add column σ6 and update low=σ2. Same low row as column σ8. Add column σ8 and reduce to zero column. Eliminated column={σ6,σ8}. A one-dimensional topological feature (cyclic structure) is born at (f=2). (vi) Column σ 10 :low(σ 10 )=σ4. Confirmed as the principal sequence. Birth / extinction pair (σ4,σ 10 Duration = 2. (vii) column σ 11 :low(σ 11 )=σ5. Same low row as column σ7. Add column σ7 and update low=σ4. Column σ 10 Same low row as above. Column σ 10 Add them together and reduce to zero. Eliminate column = {σ7,σ 10}. A one-dimensional topological feature is generated at (f=2). (viii) Column σ 12 :low(σ 12 ) = σ9. Since column σ9 has been reduced to zero, the low row σ9 is unoccupied. It is confirmed as the principal column. The cyclic structure of column σ9 is eliminated. Birth (f=2), extinction (f=3). Duration length = 1. (ix) Column σ 13 :low(σ 13 )=σ 11 . Column σ 11 Since it has been reduced to zero columns, low row σ 11 It is unoccupied. It is confirmed as the principal column. Column σ 11 This eliminates the cyclical structure. Birth (f=2), extinction (f=4). Duration = 2.

[0246] (Recording of trace tokens) From the results of the reduction process described above, trace tokens corresponding to the following two one-dimensional topological features were recorded (Figure 8(c)). Trace token A (duration = 1): Principal column identifier: σ 12 (=[v1,v2,v3], the column whose elimination has been confirmed) Low row: σ9 (=[v2,v3], column corresponding to birth) Elimination string identifier group: {σ6(=[v1,v3]),σ8(=[v1,v2])} (Columns added during the reduction process of column σ9) Trace token B (duration = 2): Principal column identifier: σ 13 (=[v3,v4,v5]) low row: σ 11 (=[v4,v5]) Elimination sequence identifier group: {σ7(=[v3,v5]),σ 10 (=[v3,v4])} (The column added during the reduction process of column σ11).

[0247] (Aggregation of contributions and identification of influential VID sets) For each trace token, the duration of the corresponding topological feature was used as a weight to assign a contribution to the VID associated with a single constituent vertex corresponding to the column identifier (original column, column corresponding to the low row, and deletion column). Column involved in trace token A (duration = 1): σ 12 ,σ9,σ6,σ8. The union of the constituent vertices is {v1, v2, v3}. →Add +1 to the contribution of each VID(v1), VID(v2), and VID(v3). Column involved in trace token B (duration = 2): σ 13 ,σ 11 ,σ7,σ 10 . The union of the constituent vertices is {v3, v4, v5}. →Add +2 to the contribution of each of VID(v3), VID(v4), and VID(v5). Aggregated results: VID(v1)=1,VID(v2)=1,VID(v3)=3,VID(v4)=2,VID(v5)=2.

[0248] VID(v3) showed the highest contribution as a topological hub involved in both topological features. When the contribution threshold was set to ≥2, the set of influencing VIDs was identified as {VID(v3), VID(v4), VID(v5)} (Figure 9). This result corresponds to a configuration in which VIDs involved in topological feature B (duration = 2), which has a longer duration, are preferentially selected, and VID(v1) and VID(v2), which are involved only in feature A (duration = 1), which has a shorter duration, are excluded (Figure 8(a)). The same result is obtained when selecting the top k=3 VIDs in terms of contribution.

[0249] <Example 2: VID mapping of inference states, extraction of trace tokens, and local recalculation in the attention mechanism of a large-scale language model> This embodiment exemplifies a Transformer-based Large-Scale Language Model (LLM), but the disclosure is not limited thereto. It is similarly applicable to any inference model (including state-space models, spread models, multi-agent inference models, etc.) that can obtain an inference state as a high-dimensional state representation and map it to a virtual node identifier (VID).

[0250] (Mapping of VID to logical address space) In this embodiment, we assume an inference task in which a Transformer-based LLM is autoregressively executed in a distributed inference environment. The inference state corresponding to the i-th token in the i-th layer of the LLM is at least the key vector K i (l) , Value vector V i (l) , and the hidden state vector H i (l) This includes the following. These inference states can be stored as a KV cache on the physical GPU memory during the inference process of the LLM. The state acquisition unit 11 abstracts these inference states from the physical GPU memory and the arrangement of compute nodes and maps them uniquely as a virtual node identifier (VID) in the logical address space. The VID may be defined as a composite identifier such as in equation (7) below.

[0251]

number

[0252] Note that the components of a VID are not limited to those listed above and may include head numbers, chunk numbers, or pipeline stage numbers depending on the configuration of the inference model. Each VID is assigned a generation tag corresponding to the updated generation of the inference, and the mapping of trace tokens is limited to references that share the same generation tag. If the generation tags do not match, the identification of the affected VID set is stopped and a safe transition is performed.

[0253] (Configuration of filtration based on inference dependency structure) The phase analysis unit 12 constructs a simplex complex with the VID group as vertices (0-simplex). The distance scale D between vertices may be defined based on the dependency relationships between tokens in the attention mechanism of the LLM. For example, in the i-th layer, the attention weight from token i to token j is 1-α (i,j) (l) Alternatively, it may be defined as shown in equation (8) below.

[0254]

number

[0255] In this case, tokens with strong semantic dependencies have smaller distances and are joined as 1-simples (edges) at an early stage (low threshold) of filtration. However, the distance scale D is not limited to a definition based on attention weights. It may be defined based on inference dependency structures (reference relationships on the inference graph, data dependency relationships on the computation graph, or cosine distance between intermediate representations, etc.), or it may be defined as a weighted combination of multiple distance indices. The topological analysis unit 12 constructs filtration based on the distance scale D and the neighbor rule R or threshold sequence, and calculates topological features by performing or incrementally updating persistent homology analysis.

[0256] (External input modification and extraction of trace tokens) Consider a scenario where, during LLM inference (for example, during an interactive session), an external input change occurs where the user modifies part of a previous input prompt. Typically, this change requires a global recalculation of the KV cache for all tokens existing from the position of the modified token onward.

[0257] However, in this embodiment, the difference determination unit 13 incrementally updates the reduction process of the boundary matrix in response to the change in the external input, and extracts trace tokens that include identifiers of principal columns involved in determining the birth, disappearance, or change of topological features during the update process, and identifiers of deleted columns that were added for the deletion process on the columns to be reduced during the reduction process.

[0258] Here, the range over which the effect of prompt modification propagates through the attention mechanism is not necessarily limited to all tokens from the modification point onward. Tokens with low attention weights for the modified token (tokens with weak semantic dependency) are located farther from the modified token on the filtration, and therefore contribute less to the change in the topological structure.

[0259] (Identifying the set of influencing VIDs and performing local recalculations) The difference determination unit 13 maps the set of columns identified by the column identifiers (identifiers of the original column and the deleted column) contained in the trace token to a single identifier, obtains the set of vertices corresponding to the single identifier, and further maps it to a set of VIDs based on mapping information that references the VIDs associated with those vertices. The difference determination unit 13 calculates the contribution of each VID by weighting the set of VIDs based on the duration of topological features, and determines the set of influential VIDs based on the contribution.

[0260] In this embodiment, the affected VID set corresponds to a specific group of KV caches that are strongly topologically affected by prompt modification. That is, the affected VID set includes KV caches of tokens that have a strong semantic dependency on the modified token and that substantially contributed to the change in topological structure.

[0261] The recalculation control unit 14 fixes (reuses) the KV cache (boundary VID) values ​​of tokens not included in the influence VID set as constants, and then instructs only the computation nodes corresponding to the influence VID set to perform local recalculations of the KV cache and hidden state vector. This omits the recalculation of the KV cache for tokens unrelated to prompt modification.

[0262] (Calculation of SavedCost and allocation of resources to explore) The reserve cost calculation unit 19 quantifies the amount of physical computing resources (for example, at least one of the number of floating-point operations, memory bandwidth consumption, or power consumption) related to the recalculation of the KV cache that was avoided by the local recalculation as SavedCost.

[0263] The resource conversion and allocation unit 15 converts the SavedCost into an increment ΔB of the number of search branches or beam width based on the resource / logical conversion coefficient η. Then, it converts and allocates ΔB to the incomplete search frontier in the LLM's autoregressive text generation (for example, the parallel search width of the beam search in subsequent token generation, or the number of branches in the tree search).

[0264] This means that the reduction in computational resources due to local recalculation is directly converted into improved search accuracy within the same inference task (searching for a wider variety of generation candidates).

[0265] (Examples of effects in this embodiment) For example, if a portion of the prompt is modified in a 1024-token dialogue session, and the affected VID set is limited to 20% of the tokens, then the KV cache recalculation for the remaining 80% of the tokens is omitted. The physical computational resources required for the omitted recalculations are quantified as SavedCost, and a portion of this is converted and allocated as an increment ΔB of the beam width, allowing more candidate paths to be explored in the generation of the modified text.

[0266] Note that the above figures (1024 tokens, 20%, etc.) are examples for the purpose of explaining this embodiment, and the size of the affected VID set in actual operation may vary depending on the modifications, the configuration of the inference model, and the attention structure.

[0267] (Notes regarding the generalization of this embodiment) In this embodiment, a configuration is shown in which the KV cache in the LLM's Attention mechanism is mapped to VID, but the scope of application of this disclosure is not limited to this. Similar local recalculation control and resource redirection allocation can be achieved if inference states are generated during the inference process, these inference states can be associated with virtual node identifiers (VID), and filtration-based phase analysis can be applied to the set of these inference states.

[0268] in particular, (a) Intermediate states in inference models other than Transformer (state-space models, diffusion models, etc.) (b) Agent state in multi-agent inference, (c) Physical quantities of grid points or mesh points in scientific simulations, (d) Variable constraints in a constraint satisfaction problem, (e) The embedding vector of the reference document chunk in Search Enhancement Generation (RAG) You can map each to a VID and apply them similarly.

[0269] <Example 3: Implementation of the processing chain of the present disclosure using the inference state of a real Transformer model, and verification of local recalculation targeting the KV cache> This embodiment shows an example of implementing and executing the processing chain described in Embodiments 1 and 2 on an actual inference state generated by a Transformer-based language model. This embodiment specifically demonstrates that the processing chain from the state acquisition unit 11 to the recalculation control unit 14 can be implemented and executed on the internal state of an actual inference model. However, the model configuration, state source, number of dimensions, various numbers, and parameters in this embodiment are all illustrative examples for explanation purposes only and are not intended to limit the scope of this disclosure.

[0270] (Part 1: Implementation of a processing chain using the hidden state as the inference state) The state acquisition unit 11 acquires a hidden state vector (896 dimensions as an example) of the final token position obtained at each generation step in the autoregressive generation process of a Transformer-based CausalLM (here, a pre-trained language model "Qwen2.5-0.5B-Instruct" with approximately 0.5B parameters), and stores this hidden state vector as an inference state to which a virtual node identifier VID is assigned. In this example implementation, 26 VID-assigned inference states were obtained from this generation process.

[0271] The phase analysis unit 12 considered the 26 VID-assigned inference states as a point cloud in a high-dimensional space and constructed a simplicial complex based on the distance scale between states. In this example implementation, the simplicial complex was obtained as a configuration containing 232 simplices. The phase analysis unit 12 constructed a boundary matrix on the Z2 coefficients corresponding to the simplicial complex and performed column reduction processing according to the order of filtration.

[0272] The difference determination unit 13 extracted trace tokens from the column reduction process, including the identifier of the source column, the identifier of the column corresponding to the low row, and the identifier of the deleted column that was added for the deletion process on the column to be reduced. In this example implementation, 206 trace tokens were recorded from the column reduction process.

[0273] The difference determination unit 13 maps the set of columns identified by the column identifiers contained in the trace token to a single identifier, obtains the constituent vertices corresponding to the single identifier, and further maps them to a set of VIDs based on the mapping information that references the VIDs associated with those vertices. The difference determination unit 13 applies weights based on duration to the set of VIDs to aggregate the contribution of each VID, and determines the set of influential VIDs based on the contribution. In this example implementation, 17 of the 26 VIDs were identified as the set of influential VIDs.

[0274] The recalculation control unit 14 generated a local recalculation instruction that targets the influence VID set for recalculation. The reserve capacity calculation unit 19 evaluated the number of states for which recalculation was avoided as 9, based on the difference between the number of states when the recalculation target is all inference states (26) and the number of states when it is limited to the influence VID set (17).

[0275] Here, the number of avoided states (9) in this example implementation is calculated as a state number proxy based on the number of inference states excluded from recalculation, and is not the SavedCost itself (at least one of the number of floating-point operations reduced by avoiding recalculation, memory bandwidth usage, communication bandwidth usage, or reduction in execution time) as defined above. This state number proxy is an indicator for easily grasping the degree of localization by the affected VID set, and SavedCost as a physical computing resource can be quantified based on the amount of data whose recalculation was avoided, as shown in Part 2 below.

[0276] Furthermore, in one example of this implementation, the distance threshold and the number of neighbors k in filtration were varied, and the results of identifying the affected VID set were evaluated. When the change in the cosine distance of the hidden states when a part of the external input was changed was used as the external proxy label for evaluation, it was confirmed that the amount of leakage and reduction in the affected VID set was sensitive to the distance threshold and the number of neighbors k. Specifically, when the distance threshold and the number of neighbors k were set to be small to construct a sparse simplicial complex (towards a narrower estimate of the influence range), the number of states included in the affected VID set decreased, resulting in a larger reduction, but it was possible to overlook the changed VIDs on the proxy. Conversely, when the distance threshold and the number of neighbors k were set to be large to construct a dense simplicial complex (towards a wider estimate of the influence range), the number of states included in the affected VID set decreased, resulting in a smaller reduction, but it was possible to include all the changed VIDs on the proxy. This trade-off was observed.

[0277] The observed results demonstrate that the configuration in which the difference determination unit 13 adjusts at least one of the threshold and k according to reliability information, reliability score, alternative reliability score, or data health (a configuration that prioritizes leak suppression as reliability decreases and efficiency as reliability increases), and the configuration of a safety-side transition that stops identifying the affected VID set and extends to wide-area recalculation when reliability falls below a predetermined reliability threshold, functions to achieve both reduction and leak suppression. In other words, the technical significance of this configuration lies in the fact that, rather than grasping the size of the affected VID set or the reduction amount itself as a fixed effect, the degree of localization can be controlled while ensuring the consistency of the inference through the decision rule (adjustment according to the threshold and upper k and reliability) and the safety-side transition.

[0278] Furthermore, in one example of this implementation, it was confirmed by reuse proxy (top-level match) that, under conservative settings, the LM head readout results do not break down when reusing the inference state corresponding to a VID that does not belong to the influential VID set (boundary VID). This confirmed that a local recalculation configuration that fixes and reuses the state of a VID located outside the influential VID set as a boundary condition can be applied while maintaining consistency.

[0279] (Part 2: Local recalculation and quantification of saved cost for KV caches) The state acquisition unit 11 acquired the KV cache (past key-value pairs) held during the Transformer-based CausalLM inference process as an inference state with VIDs, in units of token position, layer, and KV component. Each VID may be composed of a composite identifier including, for example, (RequestID, LayerID, TokenIndex).

[0280] The phase analysis unit 12 constructed a simplex complex using the VID-equipped KV cache states as a point cloud and performed column reduction on the Z2 boundary matrix. The difference determination unit 13 extracted trace tokens from the column reduction process, including the identifier of the principal column, the identifier of the low column, and the identifier of the deleted column, and identified the influential VID set by mapping the column identifiers to the VID set via the simplex and the constituent vertices. In this example implementation, 11,300 simplex states and 2,527 trace tokens (of which 1,443 trace tokens were related to the determination of the influential VID set) were recorded, and 524 KV states were identified as influential VIDs.

[0281] The recalculation control unit 14 fixed (reused) KV states that did not belong to the influential VID set as boundary conditions and instructed local recalculation limited to KV states corresponding to the influential VID set. In this example implementation, there were 364 reusable KV states. The reserve capacity calculation unit 19 calculated the amount of data related to the KV cache whose recalculation was avoided by this localization as the reduced KV data amount (approximately 186,368 bytes as an example). This reduced KV data amount was equivalent to approximately 0.41 (approximately 41%) of the amount of KV data that would require transfer or recalculation if the recalculation target was the entire cache.

[0282] Here, the reduced KV data amount and its percentage can be understood as a physical computing resource amount corresponding to the reduction in memory bandwidth usage or data transfer amount, particularly from the SavedCost as defined above, and corresponds more directly to the definition of SavedCost as a physical computing resource amount than the state number proxy shown in Part 1. The resource conversion and allocation unit 15 can convert the SavedCost into an additional number of search branches or an increment of beam width ΔB based on the resource-logic conversion coefficient η, and reallocate it to branches belonging to incomplete search frontiers within the same inference task.

[0283] Furthermore, in one example of this implementation, instead of saving all steps of the column reduction process, only the columns and rows located near the trace token are saved as a local matrix (hereinafter referred to as the local reduction window) based on the identifiers of the source column, low column, and elimination column contained in the trace token. This makes it possible to refer to the reduction process of the vicinity of the trace token actually used to determine the influential VID set as audit data recorded in the runtime output, without relying on estimation during inference execution. This configuration makes it possible to track not only the influential VID set result, but also which column operations in the boundary matrix reduction process mapped to which VID set via which individual, thereby improving the auditability of this disclosure.

[0284] Note that the numerical values ​​such as the model configuration, number of dimensions, number of states, number of units, number of trace tokens, number of influencing VIDs, amount of reduced KV data, and their percentages shown in this embodiment are all illustrative examples for the purpose of explaining this embodiment, and the values ​​in actual operation may vary depending on the inference model configuration, input content, distance scale D, nearest neighbor rule R, threshold sequence, and determination rule for the influencing VID set.

[0285] <Evaluation Example (Bench Simulation Definition)> The following are examples of benchmarks and simulations for evaluating the effectiveness of this disclosure. These are illustrative examples, and the evaluation environment, tasks, and metrics are not limited.

[0286] The evaluation environment is a distributed inference environment in which the same inference task is explored in parallel via multiple branches using an inference model distributed across multiple computing nodes (including CPUs or GPUs). Each computing node may maintain inference state sets, inference dependency structures, and correspondence information between VIDs and computing nodes. The evaluation environment may be a real cluster or a simulator with added network and node failure models, including communication delay, bandwidth constraints, and node failures. For example, the number of VIDs N may be set in the range of 10 to 10,000, the number of computing nodes in the range of 2 to 128, the number of clusters in the range of 2 to 20, and the communication delay in the range of 0 to 100 milliseconds. These values ​​are examples and are not limiting.

[0287] The evaluation task is an inference task that includes interactive inference, search extension generation, planned search, or constraint satisfaction, in which a portion of the inference state set may be invalidated or updated by a change in external input. The evaluation task may include a mode in which the search frontier is maintained within the same task and incomplete branches continuously exist.

[0288] At least the following methods may be used as comparison methods. (a) Overall recalculation method: A method that recalculates all or a predetermined wide range of inference states when an external input is changed or updated. (b) Local recalculation method (without phase analysis): A method that estimates the scope of influence based on a predetermined closure of the inference-dependent structure and performs local recalculations. (c) The disclosed method: A method that determines the influential VID set by calculating the contribution from the column identifiers of the reduced processing of the persistent homology analysis (including, if necessary, the elimination column or a generalized column set based on the origin set), performs a local recalculation limited to the influential VID set, and further converts and allocates the saved cost to the number of search branches or beam width.

[0289] During evaluation, injection events for external input changes or inference state updates may be generated, for example, in the following types: (a) Input revision type (b) External reference update type (c) Identity undetermined type (d) type with defects (e) Approximation degree change type These types are examples and may be combined.

[0290] For example, the following can be used as measurement indicators. (a) Recalculation cost: FLOPs, execution time, memory bandwidth, communication bandwidth, or a combination thereof. (b) Saved Cost (c) Exploration Enhancement Amount: The total amount of the additional number of exploration branches or the increase in beam width. (d) Success rate of confirmation: Percentage of cases that were confirmed, number of reassignments, percentage of cases that were discarded. (e) Inference quality: agreement rate, constraint satisfaction rate, consistency after reference updates, etc. (f) Leakage / overestimation of the scope of impact: Leakage rate and overestimation rate. (g) Stability: switching frequency and number of safe transitions.

[0291] You may add communication delays, bandwidth limitations, message retransmissions, duplicate receipts, and node downtime to simulate situations where allocation confirmation messages and backup / restart packages are retransmitted or duplicated.

[0292] In this disclosure method, the trade-off between reducing the omission rate in impact scope estimation and suppressing overestimation of the impact set can be controlled by switching the selection range of column identifiers (primary column only, primary column + deletion column, or generalization based on the origin set).

[0293] In this disclosure method, by using a setting that includes the deletion column in addition to the principal column as the column identifier, the omission rate of the influence scope estimation can be reduced. On the other hand, if the set of influencing VIDs becomes excessive, the deterioration of inference quality can be suppressed by re-extracting important individual elements or extending to a wide-area recalculation. Here, for the sake of reproducibility, the true value set I and the estimated set I in the evaluation may be defined as follows: That is, let S be the set of VIDs that are directly perturbed by the injection event, and let the propagation rule in the inference dependency structure be: (a) Hop-closed type, where I is the set of VIDs that can be reached within a predetermined number of hops h, And, (b) A probability propagation type in which a predetermined propagation probability is assigned to each edge, and VID set I is the set of VIDs for which the probability of arrival is greater than or equal to a threshold τ. Adopt at least one of the following: The rule and parameters (h, τ) may be fixed for evaluation.

[0294] Estimated set I is the set of influence VIDs calculated for each comparison method. Note that the evaluation of omissions and overestimations is, for example, based on the true value set I. true , the estimated set is I est Therefore, we may define FN (false negative rate) as formula (9) and FP (excess rate) as formula (10).

[0295]

number

[0296] Herein, FP is defined as the excess rate of the true value set criterion.

[0297] For the same event sequence and the same S, (i) A configuration in which only the principal column is used as the column identifier, (ii) A configuration in which the source column + deletion column is used as the column identifier, (iii) A configuration in which a generalized set of columns based on the set of origins (provenance) is used as the column identifier, You can compare them.

[0298] By converting SavedCost into the number of search branches or beam width using the coefficient η, and reallocating it only to incomplete branches, the amount of search enhancement can be increased, potentially leading to a reduction in the average execution time at the same quality level, or an improvement in the search success rate at the same time level. Here, "equivalent quality" means that the inference quality satisfies predetermined conditions.

[0299] Here, in order to make the estimation and calibration of η reproducible, the following procedure may be fixed, for example: short-time profiling during or before the start of inference to determine the additional cost c per step of the search (1 branch extension, or beam width + 1). step This can be measured or modeled as at least one of the following: execution time, FLOPs, bandwidth, etc., and its average value can be used to define it as shown in the following formula (11). The η calculated by formula (11) base This represents the physical cost C required for one step of the search (e.g., expanding the beam width by 1). step It is the reciprocal of the value, and represents the physical cost of the remaining power equivalent to the recalculations avoided by local recalculations, determining the baseline value for how many more steps the search can be deepened logically.

[0300]

number

[0301] Furthermore, the recalculation control unit 14 or the resource conversion and allocation unit 15 controls the topological entropy H topoThe gate coefficient η based on reliability indicators such as gate For example, the following formula (12) (where g is a monotonically non-increasing function or other decreasing function) may be used. The gate coefficient η gate is the "reliability brake (governance)" calculated from the topological entropy H topo which is a topological property of the latent space. When the model is confused and H topo is increasing (= low reliability), η gate approaches 0 by the decreasing function g, and ΔB described later is forced to shrink (be suppressed). [[ID=P11]] [[ID=P12]]

[0302] [[ID=P13]] [[ID=P14]]

Equation

[0303] [[ID=P21]] Here, the amount of computational resource savings is not a simple one that can be measured only by the execution time (T), but is also greatly related to the number of operations of the arithmetic unit (FLOPs: F) and the communication bandwidth (B) between nodes. For example, even if there is spare capacity in the arithmetic unit, the system cannot be accelerated in a situation where the communication bandwidth between nodes becomes a bottleneck. Therefore, the resource conversion distribution unit 15 may define SavedCost as a multi-objective (multi-objective) three-dimensional vector space as shown in formula (13). In formula (13), F saved is the reduced total number of operations (unit: FLOPs), B saved is the reduced total data transfer amount (unit: Bytes) (however, it is not just the bandwidth (bps), but the cumulative data amount that avoids occupancy), and T saved = the reduced total execution time (unit: Seconds). [[ID=P29]] [[ID=P30]]

[0304] [[ID=P31]] [[ID=P32]]

Equation

[0305] [[ID=P39]] Furthermore, in order to convert the amount of SavedCost into the exploration reinforcement amount, the resource conversion distribution unit 15 scalar-converts formula (13) with the unit conversion coefficient κ for each arithmetic resource class to obtain the following conversion value C savedUsing the mathematical formula (14), for example, the increment ΔB of the search branching number or the beam width can be calculated as a discrete integer by the following mathematical formula (15), and an unfinished branch can be selected from the search frontier and applied. The calculation formulas of ΔB described in this specification, that is, the mathematical formula (3) of [Equation 1], [Equation 6], the mathematical formula (15) of [Equation 16], and the mathematical formula (27) of [Equation 22] are all general forms

Number

[0306]

Number

[0307] Also, to clarify the effectiveness of the unfinished limit, (a) a mode of distributing limited to unfinished branches, and (b) a control mode of distributing without limitation, You may compare these using the same event sequence and measure the amount of search enhancement (total amount of ΔB), average execution time, search success rate, and confirmation success rate (percentage of results reached, number of reallocations, discard rate), etc.

[0308] (Topological entropy H topo (Examples of calculation and control of η) The recalculation control unit 14 or the resource conversion and allocation unit 15 calculates the persistence diagram or persistence pairs included in the barcode representation that are generated based on filtration. i -d i Using ), the topological entropy H topo The topological entropy H can be calculated as follows: For example, equation (16) can be used to calculate the lifetime of each phase feature, and a normalized probability distribution can be constructed using equation (17). topo You may define it as follows:

[0309]

number

[0310] Also, H topo Alternatively, an index representing the confidence of the search or recalculation may be constructed using the proportion of the sum of lifetimes of the top K features, the number of features above a predetermined lifetime threshold, or the variance of the lifetime distribution. gate This can also be given as, for example, the following formula (19), g(H topo ) is H topo It may also be constructed as a monotonically non-increasing function of .

[0311]

number

[0312] Also, g(H topo ) is H topo The function is not limited to a monotonically non-increasing function, but may also be constructed as a decreasing function including the following aspects. (a)H topoThe time series values ​​are smoothed by applying a moving average or exponential moving average with a predetermined time window width w. smooth Calculate the H smooth A mode in which a monotonically non-increasing function is applied to H. topo This allows us to suppress hypersensitivity of η to local fluctuations while maintaining a global decreasing trend. (b)H topo This embodiment divides the range of values ​​into multiple intervals and constructs it as a piecewise decreasing function that progressively reduces the value of η corresponding to each interval. In this embodiment, since the value of η takes discrete step values, the calculation result of ΔB becomes stable, and the determination in the allocation determination process can be simplified. (c)H topo This embodiment applies a reduction function to a weighted coupling with the aforementioned physical load indicator (including queue dwell time or memory bus occupancy). In this embodiment, the resource-logic conversion coefficient η can be defined as a single reduction function that integrates both topological entropy and physical load.

[0313] All of the above embodiments are functions in which the output of η tends to decrease with increasing input (topological entropy or physical load index), and correspond to the decreasing function as defined herein. The selection of embodiments depends on the characteristics of the distributed inference environment (number of nodes, range of variation in communication delay, H topo The settings may be determined according to the frequency of fluctuations, etc., and the same device may switch between modes depending on the operating conditions.

[0314] (Design guidelines for selecting a decreasing function) In selecting a decreasing function to define the resource / logical transformation coefficient η, a monotonically non-increasing function is simple to implement and has minimal computational overhead, whereas H topo η can fluctuate sharply in response to observational noise or transient outliers. Therefore, H topo In environments with a high frequency of fluctuations, or in environments where the measurement of physical load indicators involves jitter, the smoothing-based decreasing function of embodiment (a) or the piecewise decreasing function of embodiment (b) may be preferable. On the other hand, H topoIn environments where fluctuations are gradual and the measurement accuracy of the physical load index is high, a monotonically non-increasing function may be preferable from the viewpoint of responsiveness and ease of implementation. In either embodiment, it is sufficient that the function tends to decrease in output as the input increases, and the specific form of the function is not limited.

[0315] In situations where the summarization rate or distributed update delay increases and error boundary information increases, the occurrence of misallocation and inference inconsistencies can be suppressed by using a setting that stops identifying the affected VID set and extends to wide-area recalculation when the reliability information falls below a threshold. Furthermore, chattering can be suppressed by hysteresis properties. Here, in order to make the error boundary information and reliability information reproducible, for example, the error boundary information ErrBound may be defined as a monotonically increasing function based on at least one of the summarization rate r, distributed update delay amount d, and missing value rate m, as shown in the following example equation (20), and the reliability information Rel may be defined as a monotonic transformation such as equation (21) (a, b, and c are fixed values ​​or values ​​fixed by prior small-scale measurements).

[0316] ErrBound=a·r+b·d+c·m ···(20) Rel=exp(-ErrBound) ···(21) For determining the safe transition, Rel is the trigger threshold τ. on Below this threshold, it transitions to a wider area, and the return threshold τ off Return to local area with ultra (τ on <τ off Hysteresis such as ) may be applied. In the evaluation, the summarization rate r, distributed update delay amount d, and missing rate m are swept, and the number of global transitions, the inconsistency rate (e.g., the rate in which constraint violations occur), and the number of switchings (chattering index) are measured, and the presence or absence of hysteresis and the threshold difference (τ) are evaluated. off / τ on You can compare the differences in stability and quality due to these factors.

[0317] (Examples of misjudgments and release conditions: hysteresis) The recalculation control unit 14 or the resource conversion and allocation unit 15 sets a first threshold H as a condition for applying suppression or search branch width increment ΔB. low and the second threshold Hhigh (H low <H high Hysteresis using ) may be introduced. For example, H topo ≥H high If the condition is met K times consecutively (when the topological entropy is high and the reliability is low), the system transitions to a suppression state and the application of ΔB is suppressed, and in the suppression state, H topo ≤H low The system may be configured to release the suppression and allow the addition of ΔB if the condition is met K times consecutively (when the topological entropy is low and the confidence level is restored). This can suppress chattering, where the state switches frequently near the confidence level boundary.

[0318] Furthermore, the conditions for releasing the suppression state are: (i) The results of the local recalculation satisfy the specified consistency conditions, (ii) Consistency must be confirmed by an external audit system. (iii) The fallback recalculation by the von Neumann type arithmetic circuit is completed. It may include at least one of the following.

[0319] In situations where a missing event causes missing values ​​in the inference state group or inference-dependent structure, the consistency of the inference can be maintained through safe transitions or allocation degeneracy by performing missing value imputation estimation and updating error boundary information. Here, in order to verify the missing value imputation estimation with high accuracy, the missing value rate m is swept, (a) The degree of agreement of the complementary estimation (e.g., neighborhood consistency rate, constraint consistency rate, or reference consistency rate, etc.), (b) Variations in error boundary information ErrBound and reliability information Rel, (c) Number of safe transitions or allocation degeneracy, (d) Inference quality (constraint satisfaction rate, matching rate, etc.) These can be measured separately. This allows for reproducible confirmation of the behavior in which local operation is maintained when the interpolation is highly accurate, and when the uncertainty of the interpolation increases, the behavior shifts to a safe-side transition or allocation degeneracy.

[0320] It can be confirmed that duplicate application suppression based on idempotent keys can suppress recalculations and over-allocation caused by duplicate application. To make this "confirmation" valid as an evaluation, for example, a test (duplicate injection test) may be performed in which message retransmission and duplicate receipt are intentionally injected. That is, for at least one of the following, such as allocation proposals, confirmation requests, confirmation responses, cancellation notices, backup data, or restart packages, the probability p-dup (including retransmission rate, etc.) that a message with the same idempotent key is received duplicately may be swept and the following indicators may be measured. (a) OverApplyCount: The number of instances where the same idempotent key has been "applied" two or more times (ideally 0). (b) DuplicateIgnoredRate: The rate at which duplicate receipts are idempotently ignored. (c) SavedCost double-counting rate: The percentage of SavedCost or ΔB that are added twice due to the same idempotent key (ideally 0).

[0321] (supplement) In addition to idempotent keys, common header information (version number, schema identifier, time information, correlation identifier, etc.) may be added to improve interoperability and auditability. Furthermore, column identifiers, generation tags, mapping resolution results, contributions, and impact VID sets may be recorded or output as audit logs, and it may be verified that the impact VID set I can be reconstructed solely from the audit log (audit reproducibility). For example, set I reconstructed based on the log. replay Set I, determined during operation logged The degree of agreement can be measured as HashMatchRate (hash match rate of the normalized representation of the set) or the Jaccard coefficient, and by confirming that it is above a predetermined threshold, auditability can be confirmed in a practical way. Formula (22) below shows an example of how to calculate the Jaccard coefficient J.

[0322]

number

[0323] (Example of saving audit records) At least one of the audit records and audit headers output by the recalculation control unit 14 or the resource transfer allocation unit 15 may be collected by an external audit system and stored in an append-only log, a write-once-read-many (WORM) storage, or a ledger equipped with a tamper detection mechanism. The external audit system can cross-reference audit records across different nodes or different implementations using device identification information, implementation identification information, generation tags (generation_tag), or snapshot identification information to verify the consistency of the local recalculation execution history, the set of affected VIDs, and the usage of SavedCost.

[0324] (Storage of pre-verification records) The aforementioned audit record or audit log may contain the results of a preliminary match performed by an external matching or evaluation processing unit prior to or in parallel with the execution of phase analysis by this device. The preliminary match record may include at least one of the following: (a) an evaluation score or statistical index used in the preliminary match, (b) a determination result of stability or instability based on the evaluation score, (c) threshold parameters including a distance threshold, quantization bit width, or time window used in the determination, (d) identification information of the data stream or inference state that was the subject of the preliminary match, and (e) a timestamp indicating the time the preliminary match was performed. By including the preliminary match record in the audit record, the determination accuracy of the preliminary match can be verified retrospectively by comparing it with the results of the phase analysis, and the basis for calibration of the threshold parameters of the preliminary match can be observably maintained for the external audit system.

[0325] The evaluation examples in this disclosure can also be performed with implementation-independent synthetic data. For example, (a) Generate N VIDs and create feature vectors assigned to multiple clusters (e.g., a point cloud containing the cluster centers + variance for each cluster). (b) The inference dependency structure is constructed as a graph generated with high probability within clusters and low probability between clusters. (c) Changes to the external input are given as perturbations to the feature quantities and constraint satisfaction information of points belonging to a part of the cluster, and the true value set I is propagated by the influence of a predetermined number of hops h or a predetermined probability threshold τ along the dependency structure.true Constitutes, (d) Estimated set I by the disclosure method est The following are calculated and evaluated: FN (false negative rate) is given by formula (23), and FP (excess rate) is given by formula (24).

number

[0326] (Example of formula) NetGain can be defined, for example, as the following formula (25).

[0327]

number

[0328] Furthermore, for conversion to exploration resources, the additional cost c-step per exploration step may be measured by short-time profiling, the coefficient η may be defined as the following formula (26), and the number of exploration branches or beam width increment ΔB may be calculated as the following formula (27).

[0329]

number

[0330] This disclosure method constructs a filtration based on a high-dimensional state representation of the inference state set and calculates topological features using persistent homology analysis. These topological features can be used not only to identify the influencing VID set (column identifier → simplex → vertex → VID → contribution), but also as an audit metric for the consistency or structure of the inference state set. An example of auditing is shown below.

[0331] As synthetic data, (a) an example of a state group with structure and (b) an example of a state group without structure may be generated. For example, (a) may be generated as a point group corresponding to a one-dimensional cyclic structure (a point arrangement in the shape of a closed curve), and (b) may be generated as a uniform random point group of the same number of samples (or a point arrangement dominated by random perturbations). Higher-dimensional state representations are not limited to two dimensions, the dimension of the feature vector is arbitrary, and the distance scale D and the nearest neighbor rule R (or threshold sequence Θ) may be fixed.

[0332] The phase analysis unit 12 constructs filtration on the point cloud and performs persistent homology analysis to calculate topological features (Figure 5). As an example of an audit metric, the maximum duration length of birth / extinction pairs corresponding to a one-dimensional feature (cyclicity), or the statistics (mean, median, etc.) of the top k duration lengths, may be defined as the consistency score ConsScore. For example, it may be defined as ConsScore = max(PersistLenH1) (where PersistLenH1 is the duration length of the one-dimensional feature).

[0333] In a set of states with structure, features with relatively long durations of one-dimensional features are observed, and the ConsScore tends to be high. On the other hand, in a set of states lacking structure, one-dimensional features disappear quickly, and the ConsScore tends to be low. This can help determine whether a set of inference states retains a "formative" structure, or to detect signs of change in such a structure.

[0334] For reproducibility, at least the sample size N, the initial value (seed) for random number generation, the distance scale D, the nearest neighbor rule R (or threshold sequence Θ), and the definition of ConsScore (maximum duration, or top k statistics, etc.) may be fixed. Furthermore, when introducing summarization, approximation, variance updating, or switching of analysis resolution, the method for calculating error boundary information and reliability information, as well as the thresholds for determining safe transitions (initiating threshold and return threshold), may be fixed.

[0335] If the ConsScore falls outside a predetermined range, or if reliability decreases due to an increase in error boundary information, etc., a safe transition may be taken, such as discontinuing the identification of the influencing VID set and extending to a wide-area recalculation. In addition, the threshold or upper limit k in determining the influencing VID set may be adjusted to prioritize leak suppression as reliability decreases.

[0336] During the analysis process, including the aforementioned audit, column identifiers, mapping resolution results, contributions, influential VID sets, and ConsScore may be recorded as audit logs and made reconstructible later.

[0337] Although the present disclosure has been described above based on embodiments, it is easy to infer that the present disclosure is not limited in any way to the above embodiments, and that various improvements and modifications are possible without departing from the spirit of the present disclosure. For example, each embodiment and each example may be configured by modifying the embodiment by adding some or more parts of the configuration of another embodiment or example to that embodiment, or by replacing some or more parts of the configuration of that embodiment. Also, the numerical values ​​given in the above embodiments are just examples, and it is of course possible to use other numerical values. [Industrial applicability]

[0338] This disclosure can be applied to updating the inference state, local recalculation, exploration resource allocation, and governance control in a distributed inference environment, and is widely applicable to conversational AI, decision support, search augmentation and generation, planning and exploration, and large-scale inference services. [Explanation of Symbols]

[0339] 10 Information Processing Devices 11. State acquisition unit 12 Phase Analysis Unit 13 Difference judgment part 14 Recalculation Control Unit 15 Resource Conversion and Allocation Department 16 Transfer section 17 Notification Receiving Department 18 Profile Table 19. Reserve Calculation Unit 30 External state preservation device 40. Persistence Diagram 41. 0-dimensional features (connected components) 42. One-dimensional features (circular / holes) 43. Birth Value 44 Death Value 45. Persistence 46. ​​Consistency Score (ConsScore) 47 Audit Judgment Threshold (Initiating Threshold τ) on , recovery threshold τ off ) 50 boundary matrix 51 Column ID (Single ID) 52 Pivot column 53 Elimination Columns 54. Provenance set 55 Column Identifier (Column ID or Referenced Identifier) 56 Single 57 vertices 58. Correspondence information (sequential correspondence between column identifiers, individual elements, vertices, VIDs, and sets of influencing VIDs) 59 Contribution 60th Generation Number 61 Generation Tag 62 Expired (tombstone) 63. Safe-side transition (extension to wide-area recalculation)

Claims

1. An information processing device that controls an inference environment that performs inference processing using an inference model placed on computing resources, (A) A state acquisition unit that acquires the inference state generated during the progress of the inference process as a high-dimensional state representation including a virtual node identifier VID defined on a virtual logical address space, (B) A phase analysis unit that performs or incrementally updates a filtration-based phase analysis on the set of inference states to calculate a topological feature that reflects at least one of the topological characteristics of the latent space or the structural changes of the computation graph, and generates correspondence information between the inference states and individual constituent vertices corresponding to the inference states in accordance with the configuration of the filtration, (C) Difference determination unit, triggered by a change in external input to at least a part of the inference state or an update of the inference state, extracts a record unit (hereinafter referred to as a trace token) from the boundary matrix reduction process in the internal processing of the phase analysis, which includes the identifier of the principal column generated in the process of the reduction process and the identifier of the erased column which is a column added for the erasure process to the column to be reduced in the process of the reduction process, (D) A difference determination unit that maps the set of columns identified by at least the identifier of the principal column contained in the trace token to a single identifier, obtains the set of vertices corresponding to the single identifier, further maps it to a set of VIDs based on the mapping information that references the virtual node identifier VID associated with the vertex, and identifies the set of influential VIDs that contributed to the local change on the topology based on the set of VIDs, (E) A recalculation control unit that issues a local recalculation command to the computing resource corresponding to the influencing VID set, including constraints and a group of VID identifiers to be recalculated, and causes the recalculation to be performed on the local range including the influencing VID set, An information processing device characterized by comprising:

2. An information processing apparatus according to claim 1, Based on the determination of the influence VID set by the difference determination unit and the instruction for local recalculation by the recalculation control unit, a reserve cost calculation unit calculates the difference as SavedCost between the amount of physical computing resources required when performing a wide-area recalculation without using the influence VID set and the amount of physical computing resources required when performing a local recalculation based on the influence VID set. A resource conversion and allocation unit converts the SavedCost into a logical parameter (hereinafter referred to as ΔB) which is an additional number of search branches or an increment in beam width based on the resource / logical conversion coefficient η, and redistributes and applies the ΔB to parallel branches belonging to incomplete search frontiers within the same inference task. Furthermore, The information processing device is characterized in that the SavedCost is quantified as at least one of the number of floating-point operations reduced by avoiding recalculations, memory bandwidth usage, communication bandwidth usage, or the reduction in execution time.

3. In the information processing apparatus according to claim 1, The recalculation control unit generates a local control command for at least one of the inference state, intermediate representation, cache area, compute node, or communication path corresponding to the influence VID set, based on the result of identifying the influence VID set. The information processing apparatus is characterized in that the local control instruction includes at least one of the following: local recalculation, integrity verification, cache invalidation, state value fixing, output reflection temporary suspension, processing priority change, computing resource allocation change, communication bandwidth allocation change, state saving, audit identifier issuance, impact scope record generation, integrity verification hash calculation, or transition to wide-area recalculation.

4. An information processing method for controlling an inference environment that performs inference processing using an inference model deployed on computing resources, (A) A state acquisition step of acquiring the inference state generated during the progress of the inference process as a high-dimensional state representation including a virtual node identifier VID defined on a virtual logical address space, (B) A phase analysis step which involves performing or incrementally updating a filtration-based phase analysis on the set of inference states to calculate a topological feature that reflects at least one of the topological properties of the latent space or the structural changes of the computation graph, and generating correspondence information between the inference states and the individual constituent vertices corresponding to the inference states in accordance with the configuration of the filtration, (C) A difference determination step, triggered by a change in external input to at least a part of the inference state or an update of the inference state, extracts a record unit (hereinafter referred to as a trace token) from the boundary matrix reduction process in the internal processing of the phase analysis, which includes the identifier of the principal column generated in the process of the reduction process and the identifier of the erased column, which is a column added for the erasure process to the column to be reduced in the process of the reduction process. (D) A difference determination step which involves mapping the set of columns identified by at least the identifier of the principal column contained in the trace token to a single identifier, obtaining the set of vertices corresponding to the single identifier, further mapping it to a set of VIDs based on the mapping information which refers to the virtual node identifier VID associated with the vertex, and identifying the set of influential VIDs that contributed to the local change on the topology based on the set of VIDs, (E) A recalculation control step which issues a local recalculation command to the computing resource corresponding to the influencing VID set, including constraints and a group of VID identifiers to be recalculated, and causes the recalculation to be performed corresponding to the local range including the influencing VID set, An information processing method characterized by including

5. In the information processing method described in claim 4, Based on the determination of the influence VID set by the difference determination step and the instruction for local recalculation by the recalculation control step, the difference between the amount of physical computing resources required when performing a wide-area recalculation without using the influence VID set and the amount of physical computing resources required when performing a local recalculation based on the influence VID set is calculated as SavedCost. The process further includes converting the SavedCost into a logic parameter (hereinafter referred to as ΔB) which is an increment in the number of search branches or beam width based on the resource / logic conversion coefficient η, and reallocating the ΔB to parallel branches belonging to incomplete search frontiers within the same inference task. The information processing method is characterized in that the SavedCost is quantified as at least one of the number of floating-point operations reduced by avoiding recalculations, memory bandwidth usage, communication bandwidth usage, or the reduction in execution time.

6. In the information processing method described in claim 4, The recalculation control step generates a local control instruction for at least one of the inference state, intermediate representation, cache area, compute node, or communication path corresponding to the influence VID set, based on the result of identifying the influence VID set. The information processing method is characterized in that the local control instruction includes at least one of the following: local recalculation, integrity verification, cache invalidation, state value fixing, output reflection temporary suspension, processing priority change, computing resource allocation change, communication bandwidth allocation change, state saving, audit identifier issuance, impact scope record generation, integrity verification hash calculation, or transition to wide-area recalculation.

7. A program for causing a computer to execute the information processing method described in any one of claims 4 to 6.

Images