An inference efficiency optimization method of an intelligent middle platform fusion large model

Abstract

The application discloses a kind of intelligent middle station fusion big model's inference efficiency optimization method, it is related to artificial intelligence technical field, including: the sequence to be treated is converted into symbolic representation flow, the distribution characteristics of symbolic representation flow are analyzed to determine task complexity attribute, and generate the task package to be inferred;The task package to be inferred is matched with load threshold value, when meeting preset degradation condition, the execution sequence is reconstructed through logic gate, and the narrow bandwidth path in multidimensional heterogeneous inference space is called, and difference execution instruction flow is generated;Initial inference result is obtained by running difference execution instruction flow, and the semantic drift state of initial inference result is monitored relative to expected logical trajectory;When determining that semantic drift state triggers rectification mechanism, deviation point is located and compensation rescheduling instruction carrying operation snapshot is extracted;The application improves the throughput efficiency and response speed of middle station, and achieves the effect of guaranteeing the semantic high-fidelity of complex logic task output result.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for optimizing the inference efficiency of a large-scale intelligent platform fusion model. Background Technology

[0002] As the core infrastructure for enterprises to build large-scale, intensive AI capabilities, AI middleware encapsulates general-purpose technical capabilities through componentization and service-oriented approaches. Integrating large models has become an inevitable choice for improving enterprises' cognitive intelligence level. Currently, related technologies mainly accelerate inference through static methods such as model quantization, pruning, and operator fusion. For example, by performing weight bit alignment or sparsification on core computationally intensive modules such as QKV projection and FFN, and combining this with the dynamic computation graph reconstruction and variable-length request fusion scheduling mechanism in the inference execution engine, the aim is to build an efficient execution link under a specific instruction set.

[0003] However, a deep contradiction remains between the inherent computationally intensive nature of large-scale model inference and the stability and cost-effectiveness requirements of the platform. Existing technologies still have limitations in the flexibility of computing resource allocation. Specifically, existing statically compressed and fused large models often adopt a "one-size-fits-all" execution logic, making it difficult to achieve elastic scaling of computing power based on the semantic complexity of different tasks. Due to the lack of an end-to-end collaborative mechanism covering the model layer, framework layer, and infrastructure layer, when forcibly switching to a lightweight path to execute complex semantic logic, the lack of effective real-time monitoring and dynamic compensation mechanisms makes it easy for the inference results to deviate from the expected logical trajectory. This rigid inference mode, lacking fine-grained state feedback, leads to an irreconcilable mismatch between semantic expression accuracy and inference execution efficiency, limiting the optimal solution for resource scheduling in large-scale, multi-task concurrent scenarios for intelligent platforms, and making it difficult to guarantee the high semantic fidelity of the output results of complex logical tasks. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for optimizing the inference efficiency of a large-scale intelligent platform model to solve the problem that the lack of dynamic resource scheduling and accuracy compensation mechanisms makes it difficult to balance inference efficiency and semantic accuracy.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for optimizing the inference efficiency of a large-scale intelligent platform fusion model, comprising:

[0008] Identify the core related regions of the fusion model and compress non-core regions to form a multi-dimensional heterogeneous reasoning space;

[0009] The sequence to be processed is transformed into a symbolic representation stream. The distribution characteristics of the symbolic representation stream are analyzed to determine the task complexity attribute, and a task package to be reasoned is generated.

[0010] The task package to be inferred is matched with the load threshold. When the preset degradation conditions are met, the execution sequence of the logic gates is reconstructed and the narrow bandwidth path in the multi-dimensional heterogeneous inference space is called to generate a differentiated execution instruction flow.

[0011] Run differentiated execution instruction streams to obtain initial inference results and monitor the semantic drift state of the initial inference results relative to the expected logical trajectory; when the semantic drift state triggers the correction mechanism, locate the deviation point and extract the compensation rescheduling instruction carrying the running snapshot;

[0012] In response to the compensation rescheduling instruction, high-fidelity resources are invoked based on the deviation point to perform targeted supplementary inference on the runtime snapshot, the inference results are output, and the generation logic of the task package to be inferred is adjusted based on the correction feedback.

[0013] Preferably, the method for forming a multi-dimensional heterogeneous reasoning space includes:

[0014] Obtain the norm distribution and gradient sensitivity index of each layer of the fusion model, construct an importance distribution map, and divide the core related regions and non-core regions according to the relative sensitivity relationship of each layer in the importance distribution map;

[0015] Perform structured pruning based on the importance of connections in non-core regions, removing connections that are not involved in the main inference path to form a sparse connection structure;

[0016] Based on the sparse connection structure, the parameters of the core associated regions are represented by the first parameter precision level, and the parameters of the non-core regions are represented by the second parameter precision level, which is lower than the first parameter precision level. A mapping relationship between parameter precision level and computational load is established to form a multi-dimensional heterogeneous inference space.

[0017] Preferably, the step of converting the sequence to be processed into a symbolic representation stream includes:

[0018] The sequence to be processed is standardized and cleaned to remove characters that do not conform to the encoding standard. By identifying the correlation strength between characters, the sequence to be processed is decomposed into the smallest semantic units. The smallest semantic units are mapped to the corresponding integer indices and arranged in the original order to generate a symbolic representation stream.

[0019] Preferably, the analysis of symbolic representation flow distribution characteristics to determine task complexity attributes includes:

[0020] The symbolic representation stream is linearly mapped to obtain the activation distribution state, and the activation distribution state is converted into an activation probability distribution.

[0021] Based on the activation probability distribution analysis, the semantic entropy value of the symbolic representation stream is projected onto the dynamically updated task complexity distribution space, and the task complexity attribute of the symbolic representation stream is determined according to the relative deviation position of the semantic entropy value in the task complexity distribution space.

[0022] Extract the semantic anchor index of the core related region, and encapsulate the semantic anchor index, task complexity attribute and symbolic representation flow to generate the task package to be reasoned.

[0023] Preferably, matching the task package to be inferred with the load threshold includes:

[0024] Based on the task complexity attribute of the task package to be reasoned, the resource allocation requirements of the multi-dimensional heterogeneous reasoning space are mapped to generate the real-time predicted resource load value of the task package to be reasoned.

[0025] The available resource balance of the multi-dimensional heterogeneous inference space is monitored as the load threshold, and the real-time predicted resource load value is compared with the load threshold. When the real-time predicted resource load value exceeds the load threshold, the dynamic matching result is determined to be overload.

[0026] When the dynamic matching result indicates that the load is excessive and the task complexity attribute is within the preset adjustment level, the preset degradation condition is met, and the trigger signal of the refactoring logic gate execution sequence is activated.

[0027] Preferably, the method for generating differentiated execution instruction streams includes:

[0028] In response to the trigger signal of the reconstructed logic gate execution sequence, the full operator chain in the fusion model is switched to a simplified operator chain; access requests to redundant operators in non-core areas are blocked by reconstructing logic gates, and the execution priority of key operators in core related areas is established.

[0029] Based on the semantic anchor index in the task package to be reasoned, low-bit parameters at the second parameter precision level are retrieved from the multi-dimensional heterogeneous reasoning space and loaded into the sparse connection structure to construct a narrow bandwidth path with low throughput characteristics.

[0030] Based on the semantic anchor index, the symbolic representation flow in the inference task package is decomposed into features; the core feature flow corresponding to the semantic anchor index is mapped to the key operator execution path, and the remaining non-core feature flow is mapped to the narrow bandwidth path that retrieves low bit parameters.

[0031] By integrating the execution path of key operators with the narrow-bandwidth path for retrieving low-bit parameters, the execution order of computation instructions is rearranged according to the inference timing, generating a differentiated execution instruction stream consisting of high-fidelity instructions in the core region and compressed instructions in the non-core region.

[0032] Preferably, the method for monitoring the semantic drift state of the initial inference result relative to the expected logical trajectory includes:

[0033] Run the core region high-fidelity instructions and non-core region compressed instructions in the differentiated execution instruction stream, and output the initial inference results;

[0034] Based on the semantic anchor index, the theoretical distribution trend of the sequence to be processed is synchronously mapped in the hidden layer space to form the expected logical trajectory;

[0035] Extract the actual feature vector of the initial inference result, analyze the geometric distance between the actual feature vector and the expected logical trajectory in Euclidean space, and monitor the magnitude of the change in geometric distance to generate semantic drift state.

[0036] Preferably, the method for locating the deviation point and extracting the compensation rescheduling instruction carrying the runtime snapshot includes:

[0037] When the magnitude of the change in geometric distance exceeds the preset logical tolerance threshold, a correction mechanism is triggered, which backtracks the execution timeline of the differentiated execution instruction stream and locks the calculation cycle of the first sudden change in geometric distance as the deviation range.

[0038] By comparing the distribution differences between the actual feature vectors and the expected logical trajectory within the deviation interval, the position of the word with the largest deviation is determined as the deviation point, and the corresponding hierarchical index and operator number are retrieved to establish a coordinate mapping.

[0039] Based on coordinate mapping, intermediate variables and context cache at the moment of deviation triggering are extracted to form a runtime snapshot, and a compensation rescheduling instruction is generated by combining the hierarchical index.

[0040] Preferably, the targeted supplementary inference on the runtime snapshot includes:

[0041] The hierarchical index and runtime snapshot in the compensation rescheduling instruction are parsed, and high-fidelity resources of the first parameter precision level are called in the multi-dimensional heterogeneous inference space based on the hierarchical index.

[0042] The intermediate variables and context cache from the running snapshot are loaded into the computation node where the deviation point is located, and high-fidelity resources are used to start targeted supplementary inference from the deviation point location to generate a corrected feature vector;

[0043] The abnormal feature distribution in the initial inference result is corrected by using the correction feature vector, and the sequence decoding is completed and the inference result is output.

[0044] Preferably, the logic for adjusting the generation of the task package to be inferred based on correction feedback includes:

[0045] Compare the distribution of abnormal features in the corrected feature vector with the distribution of abnormal features in the initial inference result, obtain the correction feedback information, and associate the correction feedback information with the hierarchical index of the deviation point;

[0046] Based on the correction feedback information, identify the critical point of semantic entropy value that triggers semantic drift state, and adjust the partition boundary of task complexity attribute in the dynamically updated task complexity distribution space accordingly.

[0047] Based on the adjusted boundary, the judgment parameters for analyzing the distribution characteristics of symbolic representation flow are corrected, and the iterative adjustment of the logic for generating the task package to be reasoned is completed.

[0048] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the inference efficiency optimization method for the intelligent middle platform fusion large model as described in the first aspect of the present invention.

[0049] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the inference efficiency optimization method for the intelligent middleware fusion large model as described in the first aspect of the present invention.

[0050] The beneficial effects of this invention are as follows: By identifying and constructing a multi-dimensional heterogeneous inference space through parameter importance distribution, and combining it with semantic entropy value analysis based on activation probability distribution to dynamically quantify task complexity attributes, a precise mapping between computing resource requirements and heterogeneous inference space is achieved. Subsequently, when degradation conditions are met, redundant operators are blocked by reconstructing logic gates and narrow-bandwidth paths are called to generate differentiated instruction streams. At the same time, closed-loop correction of expected logic trajectory monitoring and targeted supplementary inference execution is used to intelligently schedule the inference path of large models and correct the logic trajectory in real time. This mechanism of on-demand allocation and dynamic supplementary computing not only suppresses the contradiction between excessive energy consumption of large model inference and hardware load fluctuations, but also improves the throughput efficiency and response speed of the middle platform while ensuring the semantic high fidelity of the output results of complex logic tasks. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the inference efficiency optimization method for the intelligent middleware fusion large model in this invention.

[0053] Figure 2 This is a flowchart illustrating the formation of a multi-dimensional heterogeneous reasoning space in this invention.

[0054] Figure 3 This is a flowchart of generating differentiated execution instruction streams in this invention.

[0055] Figure 4 This is a flowchart of the output reasoning results in this invention. Detailed Implementation

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0059] Reference Figure 1 , Figure 2 , Figure 3 and Figure 4 As an embodiment of the present invention, this embodiment provides a method for optimizing the inference efficiency of a large-scale intelligent platform fusion model, comprising the following steps:

[0060] The method for forming a multi-dimensional heterogeneous reasoning space includes:

[0061] Obtain the norm distribution and gradient sensitivity index of each layer of the fusion model, construct an importance distribution map, and divide the core related regions and non-core regions according to the relative sensitivity relationship of each layer in the importance distribution map.

[0062] It should be noted that the fusion large model is a large-scale pre-trained model based on deep learning; its structure consists of a deep neural network that integrates multiple Transformer hierarchical structures (including self-attention mechanisms and feedforward neural networks), and includes full weight tensors, operator topological relationships, and activation function paths; it is used to understand, generate, logically reason about, and perform multi-task semantic alignment of complex sequence data by learning from massive amounts of cross-domain data, and is the core computing engine that provides general artificial intelligence capabilities;

[0063] By obtaining the weight tensors of each weight level in the fusion model and calculating their L2 norm, the parameter magnitude distribution of each weight level in the fusion model is characterized. During backpropagation or sensitivity testing, the gradient tensors of each weight level are extracted, and the mean absolute value of the gradient tensors is calculated to determine the gradient sensitivity index. The L2 norm distribution and the gradient sensitivity index are normalized and weighted to generate an importance distribution map that reflects the contribution of each weight level in the fusion model to the inference result. The set of neurons with weights higher than the example value of 0.15 and playing a decisive role in semantic generation is classified as the core associated region, while the weight levels with weights lower than the example value of 0.15 and located on non-critical paths are classified as non-core regions.

[0064] Perform structured pruning based on the importance of connections in non-core regions, removing connections that are not involved in the main inference path to form a sparse connection structure.

[0065] It should be noted that by sorting the weight contribution values ​​of each group of neurons in the non-core region according to their numerical values, information transmission links with extremely low weight contribution values ​​(e.g., values ​​below 0.01 in the example) that did not participate in the main inference path are identified. Then, structured masking techniques are used to perform channel-level or dimension-level zeroing operations on the weight tensors in the non-core region, directly removing the corresponding redundant operator connections and feature transmission paths. By performing compressed storage mapping on the zeroed weight tensors, the storage occupation of zero-value weights is eliminated, transforming the fusion model into a sparse connection structure that retains only efficient computational paths.

[0066] Based on the sparse connection structure, the parameters of the core associated regions are represented by the first parameter precision level, and the parameters of the non-core regions are represented by the second parameter precision level, which is lower than the first parameter precision level. A mapping relationship between parameter precision level and computational load is established to form a multi-dimensional heterogeneous inference space.

[0067] Specifically, based on the sparse connection structure, the set of neurons divided into core interconnected regions is locked, and their full weight tensors are maintained in high-precision floating-point format as the first parameter precision level, for example, FP16. For the sparse connection parts in non-core regions, heterogeneous quantization technology is used to convert the corresponding weight parameters into low-bit integer format as the second parameter precision level, for example, INT4.

[0068] By statistically analyzing the hardware throughput and memory bandwidth consumed by different operators when performing calculations at the first and second parameter precision levels, a linear regression mapping relationship between parameter precision levels and computational load is established. The core related regions with precision labels, the quantized non-core regions, and the corresponding linear regression mapping relationships are logically encapsulated to form a multi-dimensional heterogeneous inference space with multi-level precision expression and load prediction capabilities.

[0069] Transforming the sequence to be processed into a symbolic representation stream includes:

[0070] The sequence to be processed is standardized and cleaned to remove characters that do not conform to the encoding standard. By identifying the correlation strength between characters, the sequence to be processed is decomposed into the smallest semantic units. The smallest semantic units are mapped to the corresponding integer indices and arranged in the original order to generate a symbolic representation stream.

[0071] Specifically, the process involves receiving the sequence to be processed from the client and performing standardized cleaning. Regular expressions are used to match and remove abnormal characters in the sequence that do not conform to the UTF-8 encoding standard of the example values. A word segmentation dictionary containing 50,000 entries of the example values ​​is invoked to perform a full scan of the sequence. The co-occurrence probability between characters is calculated to identify the strength of association between them, breaking down the sequence into the smallest semantic units composed of roots, affixes, or complete words. The unique numerical code corresponding to each smallest semantic unit in the word segmentation dictionary is queried, mapping the smallest semantic unit to its corresponding integer index. The generated integer indices are linearly arranged according to the original logical order of the sequence to generate a symbolic representation stream with discrete characteristics.

[0072] It should be noted that the word segmentation dictionary is constructed by statistically analyzing the co-occurrence frequency of characters in large-scale cross-domain corpora, defining high-frequency character combinations as the smallest semantic units and assigning them unique integer indices, and building a table structure that supports the mutual mapping between the smallest semantic units and integer indices.

[0073] Analyzing the distribution characteristics of symbolic representations to determine task complexity attributes includes:

[0074] The symbolic representation stream is linearly mapped to obtain the activation distribution state, and the activation distribution state is converted into an activation probability distribution.

[0075] Specifically, the symbolic representation stream is input into the embedding layer of the fusion model, which converts integer indices into dense vector representations. Combined with the original full weight tensor in the fusion model, a linear projection operation is performed to extract the hidden layer feature vectors of the symbolic representation stream feature distribution, thereby obtaining the activation distribution state. By applying a normalized exponential function to the activation distribution state, the values ​​of each dimension in the hidden layer feature vector are mapped to a probability interval between 0 and 1, generating the activation probability distribution representing the predicted probability distribution of the symbolic representation stream in the semantic space.

[0076] Based on the activation probability distribution analysis, the semantic entropy value of the symbolic representation stream is projected onto the dynamically updated task complexity distribution space, and the task complexity attribute of the symbolic representation stream is determined according to the relative deviation position of the semantic entropy value in the task complexity distribution space.

[0077] Specifically, the activation probability distribution is measured using the entropy calculation logic in information theory, and the uncertainty of each prediction probability is analyzed to obtain the semantic entropy value of the symbolic representation stream, which is used to quantify the information logic complexity contained in the symbolic representation stream.

[0078] The semantic entropy value is input into the dynamically updated task complexity distribution space, and the distance between the semantic entropy value and the center point of the historical entropy value stored in the dynamically updated task complexity distribution space is compared. If the semantic entropy value is less than the preset low complexity threshold (e.g., 0.35 in the example), the semantic entropy value is determined to be in the high-frequency low-value range, and the task complexity attribute of the symbolic representation flow is determined to be a low complexity scenario; if the semantic entropy value is greater than the preset high complexity threshold (e.g., 0.75 in the example), the semantic entropy value is determined to be biased towards the extremely high value distribution area, and the task complexity attribute of the symbolic representation flow is determined to be a high complexity scenario.

[0079] The dynamically updated task complexity distribution space continuously collects the semantic entropy value set generated by the multi-dimensional heterogeneous reasoning space in the historical reasoning cycle, and uses the moving average algorithm to correct the cluster center points and boundary ranges of different complexity levels in real time, so as to ensure that the dynamically updated task complexity distribution space can automatically calibrate the judgment criteria of task complexity attributes as the distribution of client requests changes.

[0080] Extract the semantic anchor index of the core related region, and encapsulate the semantic anchor index, task complexity attribute and symbolic representation flow to generate the task package to be reasoned.

[0081] It should be noted that the parameter coordinates and operator identifiers corresponding to the core associated regions in the multi-dimensional heterogeneous inference space construction stage are retrieved from the middle platform memory to form a semantic anchor index; the semantic anchor index, the determined task complexity attribute, and the original symbolic representation stream are packaged into binary according to the preset communication protocol format to generate the task package to be inferred.

[0082] Existing large-scale model inference techniques often lack flexibility when facing dynamically fluctuating system loads. Traditional optimization methods mostly employ static quantization or pruning strategies, which prevents the system from elastically adjusting to real-time resource load and the complexity of the task itself. When high-performance requirements conflict with hardware resource bottlenecks, existing architectures typically have no choice but to queue all tasks or indiscriminately reduce overall accuracy, leading to a sharp increase in response latency under high load scenarios. This "one-size-fits-all" resource allocation model ignores the differences in the accuracy sensitivity of features within a task, making it difficult to achieve optimal utilization of computing power, memory, and bandwidth while ensuring the correctness of core semantics. Therefore, this invention achieves high-fidelity priority inference of core features through dynamic load matching and operator reconstruction, as detailed below:

[0083] Matching the task package to be inferred with the load threshold includes:

[0084] Based on the task complexity attribute of the task package to be inferred, the resource allocation requirements of the multi-dimensional heterogeneous inference space are mapped to generate the real-time predicted resource load value of the task package to be inferred.

[0085] It should be noted that the task complexity attribute of the task package to be inferred is parsed, and a linear regression mapping relationship between the pre-established parameter precision level and computational load in the multi-dimensional heterogeneous inference space is retrieved based on this attribute. The linear regression mapping relationship is used to analyze the memory usage, computational overhead, and total bandwidth requirements for executing the current symbolic representation stream. The memory usage and computational overhead are then weighted and summed to output a real-time predicted resource load value representing the hardware resource consumption intensity of the task package to be inferred during execution.

[0086] The available resource balance in the multi-dimensional heterogeneous inference space is monitored as the load threshold, and the real-time predicted resource load value is compared with the load threshold. When the real-time predicted resource load value exceeds the load threshold, the dynamic matching result is determined to be overload.

[0087] Specifically, the monitoring interface of the heterogeneous resource scheduler is invoked in real time to obtain the unused video memory, remaining floating-point computing power, and available bandwidth capacity in the current hardware acceleration unit. These remaining indicators are then converted into values ​​with the same dimensions as the real-time predicted resource load value, serving as a load threshold. The real-time predicted resource load value is compared with the load threshold. If the real-time predicted resource load value is greater than the load threshold, it is determined that the current hardware resources cannot meet the requirements of high-precision full-scale inference, thus confirming the dynamic matching result as an overload state.

[0088] When the dynamic matching result indicates that the load is excessive and the task complexity attribute is within the preset adjustment level, the preset degradation condition is met, and the trigger signal of the refactoring logic gate execution sequence is activated.

[0089] Specifically, the task complexity attribute of the task package to be inferred is extracted and matched with a preset adjustment level to determine whether the current task allows for a reduction in precision to improve execution efficiency. The preset adjustment level is determined by statistically analyzing the response latency of the multi-dimensional heterogeneous inference space under different load pressures, classifying tasks into non-adjustable, limited-adjustable, and automatically adjustable levels (e.g., the automatically adjustable level). If the dynamic matching result confirms that the task is in an overloaded state and the task complexity attribute is identified as an automatically adjustable level (e.g., corresponding to a low-complexity scenario), then the preset degradation condition is met, and the trigger signal for reconstructing the logic gate execution sequence is activated.

[0090] Methods for generating differentiated execution instruction streams include:

[0091] In response to the trigger signal of reconstructing the execution sequence of logic gates, the full operator chain in the fusion model is switched to a simplified operator chain; access requests to redundant operators in non-core areas are blocked by reconstructing logic gates, and the execution priority of key operators in core related areas is established.

[0092] Specifically, in response to the trigger signal for reconstructing the execution sequence of logic gates, the underlying operator scheduling table is modified to shield computation nodes in non-core regions. Furthermore, by setting access permission bits in the memory address space, requests from execution threads to read operator weights in non-core regions are blocked, switching the full operator chain to a simplified operator chain. Simultaneously, in the task pipeline scheduler of the multi-dimensional heterogeneous inference space, higher hardware thread priorities are allocated to key operators within the core associated regions, ensuring that feature extraction in the core associated regions is prioritized even when computational resource load exceeds the load threshold.

[0093] Based on the semantic anchor index in the task package to be inferred, low-bit parameters at the second parameter precision level are retrieved from the multi-dimensional heterogeneous inference space and loaded into the sparse connection structure to construct a narrow-bandwidth path with low throughput characteristics.

[0094] Specifically, the semantic anchor index in the task package to be reasoned is parsed to accurately locate the logical address of non-core areas, and the low bit parameters at the second parameter precision level (e.g., the example value INT4) at the corresponding address are retrieved from the multi-dimensional heterogeneous reasoning space.

[0095] The retrieved low-bit parameters are loaded into the parameter buffer of the previously formed sparse connection structure, replacing the original first parameter precision level value. In the execution pipeline of the simplified operator link, independent instruction branches are opened for non-core feature flows through logical address redirection technology, and the computation instruction orchestrator is used to point the address space corresponding to the non-core region to the low-bit parameter storage area.

[0096] Configure a dedicated low-bit multiply-accumulate computation unit, such as an arithmetic logic unit that supports INT4 systolic array operations, and set it as the core of the narrow bandwidth path. By limiting the bit width of the memory access controller corresponding to this instruction branch in the underlying hardware abstraction layer, for example, limiting it from the original 128-bit width to 32-bit width, the narrow bandwidth path is constructed.

[0097] Based on the semantic anchor index, the symbolic representation flow in the inference task package is decomposed into features; the core feature flow corresponding to the semantic anchor index is mapped to the key operator execution path, and the remaining non-core feature flow is mapped to the narrow bandwidth path that retrieves low bit parameters.

[0098] Specifically, a multi-dimensional masking segmentation operation is performed on the input symbolic representation stream using semantic anchor indexes to extract high-contribution lexical vectors strongly correlated with the semantic anchor indexes, forming the core feature stream. This core feature stream is then directed to the execution path of a key operator supported by the first parameter precision level (e.g., FP16 in the example). Simultaneously, the remaining low-contribution background lexical vectors in the symbolic representation stream are classified as non-core feature streams. These non-core feature streams are mapped to narrow-bandwidth paths that retrieve low-bit parameters, achieving heterogeneous matching between different feature streams and different operator precision paths through asymmetric feature mapping techniques.

[0099] By integrating the execution path of key operators with the narrow-bandwidth path for retrieving low-bit parameters, the execution order of computation instructions is rearranged according to the inference timing, generating a differentiated execution instruction stream consisting of high-fidelity instructions in the core region and compressed instructions in the non-core region.

[0100] Specifically, the logical dependencies between the execution paths of key operators and narrow-bandwidth paths are captured by the computation instruction orchestrator, and the operator sequences of the two types of paths are merged and topologically sorted according to the inference sequence preset by the fusion big model.

[0101] For example, when the inference timing reaches the tenth Transformer level, the computation instruction orchestrator identifies that the self-attention mechanism operator in the critical operator execution path and the feedforward neural network operator in the narrow bandwidth path have clock synchronization requirements. Based on this, the high-fidelity instructions in the core area are inserted into the preceding execution queue, and the non-core area compressed instructions are arranged immediately after it.

[0102] By inserting memory synchronization instructions into the computation sequence to coordinate the timing of data fusion at different precisions, the instruction set is finally encapsulated into a task chain with pipelined parallel characteristics, generating a differentiated execution instruction stream that reflects the physical execution logic of multi-dimensional heterogeneous inference space.

[0103] By designing reconfigurable logic gates and differentiated instruction streams for overload conditions, a leap from static resource stacking to dynamic logic reshaping has been achieved. In extreme cases where hardware resources reach their thresholds, operator link switching can be completed precisely within microseconds. By suppressing non-core areas in narrow bandwidth paths, bus bandwidth pressure and memory usage are significantly alleviated. This asymmetric feature mapping mechanism ensures that the limited FP16 high-fidelity computing power always serves the core semantic anchor point, thereby significantly reducing system throughput latency under high load scenarios without service interruption. While maintaining the stability of the inference result logic trajectory, it also improves the platform system's capacity to handle concurrent tasks and the flexibility of resource allocation.

[0104] Methods for monitoring the semantic drift of the initial inference result relative to the expected logical trajectory include:

[0105] Run the core region high-fidelity instructions and non-core region compressed instructions in the differentiated execution instruction stream, and output the initial inference results.

[0106] Specifically, differentiated instruction streams are loaded into hardware execution units managed by a heterogeneous resource scheduler. High-fidelity instructions in the core region are initiated in parallel to invoke the first precision level parameter, while compressed instructions in the non-core region are executed synchronously to invoke the second precision level parameter. During inference, the key operator execution path processes the core feature stream, while the narrow-bandwidth path processes the non-core feature stream. After feature fusion at various levels, the two feature streams are finally converted into an initial inference result containing semantic probability distributions by the decoding output layer of the fused large model.

[0107] The hardware execution unit managed by the heterogeneous resource scheduler refers to the underlying physical computing cluster that carries the operation of multi-dimensional heterogeneous inference space. It integrates graphics processors, tensor processors, and low-power operator acceleration circuits that support different computing power requirements. The heterogeneous resource scheduler uses a real-time instruction distribution protocol to assign operator operations of different precision in the differentiated execution instruction stream to hardware computing nodes with corresponding processing capabilities, thereby achieving physical layer matching between hardware throughput and the precision of differentiated execution instruction streams.

[0108] Based on the semantic anchor index, the theoretical distribution trend of the sequence to be processed is synchronously mapped in the hidden layer space to form the expected logical trajectory.

[0109] Specifically, by utilizing the semantic anchor indexes in the task package to be inferred, the core neuron nodes in the fusion model that have not undergone quantization compression and pruning are located, and the symbolic representation stream after the sequence to be processed is synchronously input into a set of parallel reference computation paths. Within the hidden layer space of each Transformer hierarchy, the reference computation paths predict the feature mean distribution of the symbolic representation stream in each clock cycle based on full high-precision weights, and construct the expected logical trajectory.

[0110] Extract the actual feature vector of the initial inference result, analyze the geometric distance between the actual feature vector and the expected logical trajectory in Euclidean space, and monitor the magnitude of the change in geometric distance to generate semantic drift state.

[0111] Specifically, the actual feature vector representing the initial inference result is extracted from the last hidden layer node of the fusion model and projected into the same Euclidean coordinate system as the expected logical trajectory.

[0112] By calculating the sum of squares of the differences between the actual feature vector and the expected logical trajectory at the corresponding computation cycle positions in each dimension and taking the square root, the geometric distance of the distribution deviation between the two is obtained. Dynamic fluctuation data of the geometric distance in the inference time sequence are continuously acquired, and semantic drift states are generated based on the amplitude of the change in geometric distance.

[0113] In current heterogeneous quantization or pruning inference techniques, once a logical deviation occurs in the output due to decreased accuracy, the system typically lacks effective real-time monitoring and repair methods. This often necessitates re-inference of the entire sequence or being forced to accept low-quality output results. This "either / or" open-loop mechanism not only wastes computational power but also limits the application of low-bit quantization techniques in demanding semantic scenarios. Therefore, this solution introduces runtime snapshot capture and local bias correction techniques to establish a low-cost fault-tolerant circuit breaker mechanism, as detailed below:

[0114] Methods for locating deviation points and extracting compensatory rescheduling instructions carrying runtime snapshots include:

[0115] When the magnitude of the change in geometric distance exceeds the preset logical tolerance threshold, a correction mechanism is triggered, which backtracks the execution timeline of the differentiated execution instruction stream and locks the calculation cycle of the first sudden change in geometric distance as the deviation range.

[0116] Specifically, the magnitude of the geometric distance change calculated during the generation of semantic drift states is acquired in real time and compared with a preset logical tolerance threshold (e.g., 0.45 in this example). When the magnitude of the geometric distance change exceeds the preset logical tolerance threshold at a specific time point, a correction mechanism is immediately triggered. This mechanism retrieves runtime log records from the multi-dimensional heterogeneous inference space and traces the runtime timeline of the differentiated execution instruction streams in the hardware execution unit. By comparing the growth slope of the geometric distance within each computation clock cycle, the specific computation cycle in which the geometric distance value first experiences a non-linear surge and deviates from the distribution range of the reference computation path is determined as the deviation interval.

[0117] By comparing the distribution differences between the actual feature vectors and the expected logical trajectory within the deviation interval, the position of the word with the largest deviation is determined as the deviation point, and the corresponding hierarchical index and operator number are retrieved to establish a coordinate mapping.

[0118] Specifically, the difference between the actual feature vector and the expected logical trajectory within the deviation interval is calculated to quantify the distribution difference, and the point of maximum difference is marked as the deviation point. Based on the temporal location of the deviation point, the corresponding hierarchical index (e.g., the twelfth layer) and operator number of the fusion model are retrieved from the middle platform memory to establish a coordinate mapping between the deviation point and the physical node.

[0119] Based on coordinate mapping, intermediate variables and context cache at the moment of deviation triggering are extracted to form a runtime snapshot, and a compensation rescheduling instruction is generated by combining the hierarchical index.

[0120] Specifically, based on coordinate mapping, the system locates the moment of deviation triggering, extracts the intermediate variable tensor, attention mechanism mask, and context cache data stored at that moment from the dynamic cache of the multi-dimensional heterogeneous inference space, encapsulates them into a runtime snapshot representing the underlying computation state, so as to preserve the underlying computation state when the deviation occurs, logically associates the runtime snapshot with the hierarchical index to be corrected, adds a priority scheduling flag, and generates a compensation rescheduling instruction that indicates the intervention of high-fidelity resources.

[0121] Performing targeted supplemental inference on runtime snapshots includes:

[0122] The hierarchical index and runtime snapshot in the compensation rescheduling instruction are parsed, and high-fidelity resources of the first parameter precision level are called in the multi-dimensional heterogeneous inference space based on the hierarchical index.

[0123] Specifically, upon receiving the compensation rescheduling instruction, the hierarchical index is extracted through the instruction parsing engine to precisely locate the specific position where semantic deviation occurs in the fused large model. Based on the hierarchical index, the low-bit quantization constraint of the corresponding level is removed from the multi-dimensional heterogeneous inference space, and the full weight tensor and key operators at the first parameter precision level (e.g., FP16 in the example) are reactivated. The heterogeneous resource scheduler requests computing power resources with high-precision floating-point operation capabilities from the hardware execution unit, thereby completing the mounting and calling of high-fidelity resources and providing accuracy assurance for subsequent targeted supplementary inference.

[0124] The intermediate variables and context cache from the running snapshot are loaded into the computation node where the deviation point is located, and high-fidelity resources are used to initiate targeted supplementary inference from the deviation point location to generate a corrected feature vector.

[0125] Specifically, the runtime snapshot is deserialized, and intermediate variable tensors, attention mechanism masks, and context cache data are extracted and precisely loaded into the computation node register where the deviation point is determined by coordinate mapping. Using the loaded runtime snapshot data as the starting state, the high-fidelity resources with the first parameter precision level are driven to reproject and transform the damaged features within the deviation range. By reconstructing the correlation strength between lexical units in the high-precision semantic space, computational errors caused by low-bit paths are eliminated, generating a corrected feature vector that reflects accurate semantic logic.

[0126] The abnormal feature distribution in the initial inference result is corrected by using the correction feature vector, and the sequence decoding is completed and the inference result is output.

[0127] Specifically, the corrected feature vector is fed back to the feature fusion layer of the fusion model, replacing the anomalous feature distribution generated by the differentiated execution instruction stream in the initial inference result through a numerical overlay operation. The corrected feature distribution is then used to re-drive the subsequent Transformer hierarchical structure for residual connections and normalization processing, ensuring that the entire semantic generation chain recovers to the expected logical trajectory. The corrected global features are then probabilistically sampled and symbolically inversely mapped through the decoding output layer of the fusion model, ultimately completing sequence decoding and outputting the inference result.

[0128] By establishing a high-precision hot-repair capability, the physical nodes where semantic drift occurs are precisely located through snapshot execution. By performing targeted supplementary reasoning with high-fidelity resources only on the deviation intervals, redundant recalculation of all operators is avoided, recovering the damaged feature distribution with minimal computational compensation. This design ensures that the multi-dimensional heterogeneous reasoning space, while pursuing extreme throughput, still possesses a closed-loop guarantee capability to align with the expected logical trajectory, achieving a highly efficient balance between reasoning accuracy and execution efficiency.

[0129] The logic for adjusting the generation of the task package to be inferred based on corrective feedback includes:

[0130] The distribution of abnormal features in the corrected feature vector is compared with that in the initial inference result to obtain the correction feedback information, and the correction feedback information is associated with the hierarchical index of the deviation point.

[0131] Specifically, the correction feature vector is compared with the corresponding anomalous feature distribution in the initial inference result using element-wise difference calculation. The residual data representing the degree and direction of feature distortion is extracted as correction feedback information. Using the hierarchical index determined when establishing the coordinate mapping, the correction feedback information is logically bound to the specific Transformer hierarchical structure where the deviation occurred, in order to determine the sensitivity difference of the fusion model to parameter accuracy at different depth levels.

[0132] Based on the correction feedback information, identify the semantic entropy value critical point that triggers the semantic drift state, and adjust the partition boundary of the task complexity attribute in the dynamically updated task complexity distribution space accordingly.

[0133] Specifically, the process involves backtracking the task packages that generate anomalous feature distributions, analyzing their corresponding initial semantic entropy values, and combining this with correction feedback information to analyze the correlation between semantic entropy values ​​and the magnitude of geometric distance changes. The minimum semantic entropy value that causes the geometric distance to exceed a preset logical tolerance threshold is defined as the semantic entropy value critical point. Using a moving average algorithm, the high complexity threshold or low complexity threshold in the dynamically updated task complexity distribution space is adjusted in real time based on the semantic entropy value critical point (e.g., lowering the low complexity threshold from the example value of 0.35 to the example value of 0.30), thereby dynamically shrinking or expanding the range for judging task complexity attributes.

[0134] Based on the adjusted boundary, the judgment parameters for analyzing the distribution characteristics of symbolic representation flow are corrected, and the iterative adjustment of the logic for generating the task package to be reasoned is completed.

[0135] Specifically, the threshold addressing table in the dynamically updated task complexity distribution space is rewritten using the adjusted partition boundaries, and the original decision parameters are covered by the corrected preset low complexity threshold and preset high complexity threshold. During subsequent processing of the sequence to be processed, the updated decision parameters are used to perform interval matching on the generated semantic entropy value, redetermining the task complexity attribute of the symbolic representation flow; this attribute is then encapsulated in the task package to be inferred and passed to the load matching stage. By changing the comparison result between the real-time predicted resource load value and the load threshold, the trigger signal for reconstructing the logic gate execution sequence is blocked from being sent to the simplified operator link.

[0136] This embodiment also provides a computer device applicable to the inference efficiency optimization method of the intelligent middleware fusion large model, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the inference efficiency optimization method of the intelligent middleware fusion large model proposed in the above embodiment.

[0137] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0138] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the inference efficiency optimization method for realizing the intelligent middleware fusion model proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0139] In summary, this invention achieves a precise mapping between computational resource requirements and the heterogeneous inference space by: constructing a multi-dimensional heterogeneous inference space through parameter importance distribution identification, and dynamically quantifying task complexity attributes by combining semantic entropy value analysis based on activation probability distribution; subsequently, when degradation conditions are met, redundant operators are blocked by reconstructing logic gates and generating differentiated instruction streams by calling narrow-bandwidth paths, while closed-loop correction of expected logic trajectory monitoring and targeted supplementary inference execution is used to intelligently schedule the inference path of large models and correct the logic trajectory in real time; this on-demand allocation and dynamic supplementary computation mechanism not only suppresses the contradiction between excessive energy consumption of large model inference and hardware load fluctuations, but also improves the throughput efficiency and response speed of the middle platform while ensuring the semantic high fidelity of the output results of complex logic tasks.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for optimizing the inference efficiency of a large-scale intelligent platform model, characterized in that, include: Identify the core related regions of the fusion model and compress non-core regions to form a multi-dimensional heterogeneous reasoning space; The sequence to be processed is transformed into a symbolic representation stream. The distribution characteristics of the symbolic representation stream are analyzed to determine the task complexity attribute, and a task package to be reasoned is generated. The task package to be inferred is matched with the load threshold. When the preset degradation conditions are met, the execution sequence of the logic gates is reconstructed and the narrow bandwidth path in the multi-dimensional heterogeneous inference space is called to generate a differentiated execution instruction flow. Run differentiated execution instruction streams to obtain initial inference results and monitor the semantic drift state of the initial inference results relative to the expected logical trajectory; when the semantic drift state triggers the correction mechanism, locate the deviation point and extract the compensation rescheduling instruction carrying the running snapshot; In response to the compensation rescheduling instruction, high-fidelity resources are invoked based on the deviation point to perform targeted supplementary inference on the runtime snapshot, the inference results are output, and the generation logic of the task package to be inferred is adjusted based on the correction feedback.

2. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 1, characterized in that, The method for forming a multi-dimensional heterogeneous reasoning space includes: Obtain the norm distribution and gradient sensitivity index of each layer of the fusion model, construct an importance distribution map, and divide the core related regions and non-core regions according to the relative sensitivity relationship of each layer in the importance distribution map; Perform structured pruning based on the importance of connections in non-core regions, removing connections that are not involved in the main inference path to form a sparse connection structure; Based on the sparse connection structure, the parameters of the core associated regions are represented by the first parameter precision level, and the parameters of the non-core regions are represented by the second parameter precision level, which is lower than the first parameter precision level. A mapping relationship between parameter precision level and computational load is established to form a multi-dimensional heterogeneous inference space.

3. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 2, characterized in that, The process of converting the sequence to be processed into a symbolic representation stream includes: The sequence to be processed is standardized and cleaned to remove characters that do not conform to the encoding standard. By identifying the correlation strength between characters, the sequence to be processed is decomposed into the smallest semantic units. The smallest semantic units are mapped to the corresponding integer indices and arranged in the original order to generate a symbolic representation stream.

4. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 3, characterized in that, The analysis of symbolic representation of flow distribution characteristics to determine task complexity attributes includes: The symbolic representation stream is linearly mapped to obtain the activation distribution state, and the activation distribution state is converted into an activation probability distribution. Based on the activation probability distribution analysis, the semantic entropy value of the symbolic representation stream is projected onto the dynamically updated task complexity distribution space, and the task complexity attribute of the symbolic representation stream is determined according to the relative deviation position of the semantic entropy value in the task complexity distribution space. Extract the semantic anchor index of the core related regions, and encapsulate the semantic anchor index, task complexity attribute and symbolic representation flow to generate the task package to be reasoned.

5. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 4, characterized in that, The step of matching the task package to be inferred with the load threshold includes: Based on the task complexity attribute of the task package to be reasoned, the resource allocation requirements of the multi-dimensional heterogeneous reasoning space are mapped to generate the real-time predicted resource load value of the task package to be reasoned. The available resource balance of the multi-dimensional heterogeneous inference space is monitored as the load threshold, and the real-time predicted resource load value is compared with the load threshold. When the real-time predicted resource load value exceeds the load threshold, the dynamic matching result is determined to be overload. When the dynamic matching result indicates that the load is excessive and the task complexity attribute is within the preset adjustment level, the preset degradation condition is met, and the trigger signal of the refactoring logic gate execution sequence is activated.

6. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 5, characterized in that, The method for generating differentiated execution instruction streams includes: In response to the trigger signal of the reconstructed logic gate execution sequence, the full operator chain in the fusion model is switched to a simplified operator chain; access requests to redundant operators in non-core areas are blocked by reconstructing logic gates, and the execution priority of key operators in core related areas is established. Based on the semantic anchor index in the task package to be reasoned, low-bit parameters at the second parameter precision level are retrieved from the multi-dimensional heterogeneous reasoning space and loaded into the sparse connection structure to construct a narrow bandwidth path with low throughput characteristics. Based on the semantic anchor index, the symbolic representation flow in the inference task package is decomposed into features; the core feature flow corresponding to the semantic anchor index is mapped to the key operator execution path, and the remaining non-core feature flow is mapped to the narrow bandwidth path that retrieves low bit parameters. By integrating the execution path of key operators with the narrow-bandwidth path for retrieving low-bit parameters, the execution order of computation instructions is rearranged according to the inference timing, generating a differentiated execution instruction stream consisting of high-fidelity instructions in the core region and compressed instructions in the non-core region.

7. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 6, characterized in that, The method for monitoring the semantic drift state of the initial inference result relative to the expected logical trajectory includes: Run the core region high-fidelity instructions and non-core region compressed instructions in the differentiated execution instruction stream, and output the initial inference results; Based on the semantic anchor index, the theoretical distribution trend of the sequence to be processed is synchronously mapped in the hidden layer space to form the expected logical trajectory; Extract the actual feature vector of the initial inference result, analyze the geometric distance between the actual feature vector and the expected logical trajectory in Euclidean space, and monitor the magnitude of the change in geometric distance to generate semantic drift state.

8. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 7, characterized in that, The method for locating the deviation point and extracting the compensation rescheduling instruction carrying the runtime snapshot includes: When the magnitude of the change in geometric distance exceeds the preset logical tolerance threshold, a correction mechanism is triggered, which backtracks the execution timeline of the differentiated execution instruction stream and locks the calculation cycle of the first sudden change in geometric distance as the deviation range. By comparing the distribution differences between the actual feature vectors and the expected logical trajectory within the deviation interval, the position of the word with the largest deviation is determined as the deviation point, and the corresponding hierarchical index and operator number are retrieved to establish a coordinate mapping. Based on coordinate mapping, intermediate variables and context cache at the moment of deviation triggering are extracted to form a runtime snapshot, and a compensation rescheduling instruction is generated by combining the hierarchical index.

9. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 8, characterized in that, The targeted supplementary reasoning for the runtime snapshot includes: The hierarchical index and runtime snapshot in the compensation rescheduling instruction are parsed, and high-fidelity resources of the first parameter precision level are called in the multi-dimensional heterogeneous inference space based on the hierarchical index. The intermediate variables and context cache from the running snapshot are loaded into the computation node where the deviation point is located, and high-fidelity resources are used to start targeted supplementary inference from the deviation point location to generate a corrected feature vector; The abnormal feature distribution in the initial inference result is corrected by using the correction feature vector, and the sequence decoding is completed and the inference result is output.

10. The method for optimizing the inference efficiency of a large-scale intelligent platform fusion model as described in claim 9, characterized in that, The logic for adjusting the generation of the task package to be inferred based on corrective feedback includes: The distribution of abnormal features in the correction feature vector is compared with that in the initial inference result to obtain correction feedback information, and the correction feedback information is associated with the hierarchical index of the deviation point; Based on the correction feedback information, identify the semantic entropy value critical point that triggers the semantic drift state, and adjust the partition boundary for the task complexity attribute in the dynamically updated task complexity distribution space accordingly. Based on the adjusted boundary, the judgment parameters for analyzing the distribution characteristics of symbolic representation flow are corrected, and the iterative adjustment of the logic for generating the task package to be reasoned is completed.

Images