563 results about "Memory footprint" patented technology

The invention provides a target recognition model reasoning optimization method and device, and the method comprises the steps: firstly carrying out the structural analysis and sensitivity evaluation of a pre-training model, extracting the structural features of each network layer, activating the distribution features, carrying out the quantitative sensitivity scoring, and constructing a data set reflecting the hierarchical features and fault-tolerant capability; and querying a quantitative configuration knowledge base based on the data set to generate a heterogeneous quantitative strategy. Layered low-bit quantization is executed according to the strategy, and a layered weighted loss function is introduced to carry out quantization perception training, so that precision loss caused by bit width compression is effectively compensated. According to the method, through hierarchical heterogeneous quantification, the model recognition precision is preserved to the maximum extent while high compression ratio and reasoning acceleration are achieved, and particularly, the performance of a high-sensitivity layer is protected. The generated heterogeneous quantitative model remarkably reduces memory occupation and power consumption, is suitable for an edge hardware platform with limited resources, forms a set of complete automatic process from analysis and configuration to training compensation, and has good universality and engineering practical value.

The invention discloses a retraining-free pruning and recombination method for a sparse expert hybrid large model, and belongs to the technical field of large model compression and optimization. The method aims at solving the problems that due to the fact that an existing sparse expert hybrid (SMoE) model needs to load all expert parameters, memory occupation is too high, and deployment is difficult. According to the method, firstly, redundant experts are identified and pruned based on routing activation statistics; then, decomposing the pruned experts into neuron-level functional fragments, and redistributing the fragments to the reserved experts according to structural similarity; and finally, original fragments and newly distributed fragments are merged in the reserved experts through a weighted clustering algorithm, so that compact experts with fewer parameters and stronger expression ability are reconstructed. According to the method, fine-grained operation is carried out at the neuron level, the inherent representation conflict and dislocation problems among experts are effectively solved, the performance of the compressed model is remarkably improved, and reliable technical support is provided for deploying a large-scale SMoE model.

The invention relates to the technical field of three-dimensional models, in particular to a drawing method for making a three-dimensional model based on three-dimensional laser point clouds, which comprises the following steps of: sequentially reading point cloud data files of the three-dimensional laser point clouds, inserting coordinate values of each point into an octree structure, and when octree node buffer areas in a memory are full, drawing a three-dimensional laser point cloud into the octree structure; and writing the node data in the buffer area into a hard disk, and traversing the hard disk octree node set from bottom to top. According to the method, the three-dimensional laser point cloud data is inserted step by step, and the hierarchical aggregated octree structure is utilized, so that the point cloud data processing efficiency is effectively improved, and the memory occupation pressure is reduced; representative points in an octree structure are fused with an average normal to construct a continuous function field, a grid model is adaptively generated in a multi-detail-level mode, and the detail retaining capacity and drawing precision of the model are improved. Furthermore, accurate identification and positioning of topological features are realized by adopting calculation of a cell complex sequence and a topological feature noise rank.

The invention provides a computing power service dynamic resource allocation method and system applied to AI model training. The method comprises the following steps: firstly, collecting real-time computing power resource use data (including computing node load, memory occupation and data transmission delay) and model training state data (including training task stage identification, model parameter updating frequency and training data batch processing progress) in AI model training; generating a computing power resource demand association feature set, constructing a computing power resource dynamic allocation decision model including resource allocation priority judgment, adjustment amplitude calculation and scheduling opportunity selection units based on the set, and outputting a computing power resource allocation scheme (including computing node number, memory capacity and data transmission bandwidth adjustment instructions) through the model. Resources are scheduled according to the scheme, and new data are collected to update the feature set, so that dynamic and accurate allocation of computing power resources is realized, and the resource utilization rate and the training efficiency are improved.

The invention provides a main Monitor service pressure optimization method in a distributed storage cluster, and the method comprises the steps: S1, collecting the load state information of each Monitor node in real time through a monitoring module, and enabling the load state information to comprise the CPU utilization rate, the memory occupancy rate and the network bandwidth occupancy rate; s2, a load balancing module calculates the load weight of each Monitor node based on the load state information, and dynamically adjusts a message distribution strategy according to the load weight; s3, the multi-stage message distribution module distributes the state updating message to a plurality of Monitor nodes for parallel processing according to the message distribution strategy; and S4, the state synchronization module maintains the state consistency among the Monitor nodes through a consistency protocol, and regularly verifies the data version of each node. According to the main Monitor service pressure optimization method in the distributed storage cluster provided by the embodiment of the invention, the load pressure of the main Monitor node in the distributed storage upgrading process is effectively reduced, and the system message processing efficiency and stability are improved.

The invention provides a computing resource scheduling method and device, electronic equipment and a storage medium, and the method comprises the steps that a computing task request issued by a user is received, and the computing task request comprises a computing unit mask parameter; a scheduling instruction is generated according to the calculation task request, the scheduling instruction is sent to a command processing module, and the scheduling instruction comprises the calculation unit mask parameters; and analyzing the calculation unit mask parameter in the scheduling instruction by a command processing module, and distributing a calculation task to a target calculation unit set according to an analysis result. According to the method and the device, static binding of masks and task flows / queues in a traditional scheme is replaced by dynamic integration of the mask parameters and the scheduling instructions, the situation that independent task flows / queues are created for each mask combination is avoided, and memory occupation and system management overhead can be remarkably reduced.

A CXL memory fault tolerance method, and a server system, a storage medium and an electronic device. The method comprises: acquiring parameter values of a group of operating parameters of CXL memory devices in a CXL memory device group, wherein the group of operating parameters are used for representing the operating states of the corresponding CXL memory devices; on the basis of the acquired parameter values of the group of operating parameters, predicting the operating states of the CXL memory devices in the CXL memory device group; and when it is predicted that there is an abnormal memory device operating abnormally in the CXL memory device group, performing controlling to execute a migration operation on memory data in the abnormal memory device, so as to migrate the memory data in the abnormal memory device to a target memory device operating normally in the CXL memory device group. By means of the present application, the problem of CXL memory fault tolerance methods in the prior art of the memory utilization rate of a server being low due to a hot standby memory occupying a server slot is solved.

The invention discloses an optimization method and device for an embedded linked list container, electronic equipment and a storage medium, and relates to the technical field of computers, the method comprises the steps that a linked list node is embedded into a host data structure to serve as a member variable, extra memory occupation of a pointer domain in a traditional linked list is omitted, and the optimization efficiency is improved. Physical storage of the nodes and the host data structure is continuous, so that cache missing can be reduced; meanwhile, the stability of the linked list under high concurrency is guaranteed through atomized insertion and deletion operations, and a traditional independent node structure and a non-atomized operation are not adopted; the technical problems that in the prior art, a traditional linked list node pointer domain occupies an extra memory, so that expenditure is increased, discontinuous node physical distribution causes cache missing, access delay is increased, and system abnormal safety and operation reliability are affected by iterator failure under high concurrency can be solved. And the technical effects of reducing the memory overhead, reducing the access delay and improving the abnormal safety and the operation reliability of the system are achieved.

The embodiment of the invention provides a dynamic resource scheduling method and system based on reinforcement learning, and the system comprises a state sensing module which is used for collecting and preprocessing the resource state data of each node in a cluster, and the resource state data at least comprises a CPU utilization rate, a memory occupancy rate, a network bandwidth utilization rate, a task queue length and a node load; the action decision module is used for outputting a scheduling action according to the current state representation vector, and the reward calculation module is used for calculating a reward value according to an actual operation result of the system; the model training module is used for training the strategy model by using a reinforcement learning algorithm, and the deployment optimization module is used for deploying the trained strategy model to a production environment. Complicated and changeable workloads and resource states in the cloud environment can be automatically dealt with, the manual intervention cost is remarkably reduced, and the scheduling efficiency and accuracy are improved. And a comprehensive and unified environment perception capability can be constructed.

A method of serving a generative transformer model includes determining a batch size to use in processing inference requests and allocating at least one prompt pipeline and at least on token pipeline to the generative transformer model to process the batch of inference requests. The number of prompt pipelines and the number of token pipelines, and the depths of the pipelines are determined based on the batch size, an average prompt length, a cache requirement per stage, and a memory footprint of model weights for the generative model per stage using a resource allocator component of the model serving system. Cache streaming is used to stream prompt cache from prompt pipelines to token pipelines to generate tokens. Cache streaming involves gather-copy operations which may be performed using compute kernels.

The invention belongs to the technical field of integrated circuit simulation verification, particularly relates to an RISC-V simulation resource dynamic generation method and system, and solves the problems of instruction consistency and variable-length instruction truncation through a metadata double-table structure and a cross-boundary instruction splicing mechanism. Virtualized two-stage translation support and abnormal injection are realized through a recursive multiple hit detection and dynamic attribute bit modification mechanism. According to the method, complete storage is replaced with lightweight metadata, memory occupation is remarkably reduced, the problem that memory occupation is linearly increased along with time is solved, simulation efficiency and consistency are improved, and the method is suitable for full-system verification of a high-performance RISC-V processor.

The invention discloses an audio test data processing method, system and device based on zero-copy NIO and a dynamic sliding window, and the method comprises the following steps: S1, building a memory mapping file channel, and directly mapping an audio file to a process address space; s2, constructing a double-buffer processing pipeline, and realizing asynchronous decoupling of a collection thread and a calculation thread by adopting a producer-consumer mode; and S3, a dynamic sliding window mechanism is adopted to maintain a real-time audio data interval, and the window size is dynamically adjusted according to a television CPU load. According to the technical scheme of the invention, the problems of high memory occupation, high detection delay and low cross-application data reading efficiency existing in audio testing of a television system in a multi-application concurrent scene are solved.

The invention relates to a lightweight neural network target detection model optimization method for an embedded device, and belongs to the technical field of target detection, and the method comprises the steps: determining the hardware constraint and detection task boundary of the embedded device; adjusting a network infrastructure based on hardware and task characteristics; network parameter redundancy is eliminated; designing a scene adaptive dynamic feature selection mechanism; compressing parameter storage precision; optimizing calculation-intensive operation; adjusting a memory access mode; an embedded deployment framework is adapted; and constructing a closed-loop iterative optimization process. The method is beneficial for realizing lightweight and efficient operation of the model, improving detection precision and stability, reducing memory occupation and power consumption, and meeting hardware constraints of embedded equipment.

The invention discloses an interactive document editing canvas method, which comprises the following steps of: establishing a data-view mapping relation directly driven by a Vue2 responsive system between a canvas container and an element array, any addition, deletion and modification operation aiming at the node description data set can be instantly and accurately reflected to the style and attribute of the corresponding DOM child node without manual intervention, so that element positioning, rendering and follow-up editing are completed in the same technical action, the delay and inconsistency risk caused by frequent and direct operation of a real DOM in a traditional scheme is eliminated, and the operation efficiency is improved. The interaction fluency, the layout intuition and the system maintainability are obviously improved; by means of a virtual DOM mechanism of Vue2, batch diff is carried out and then submitted to a real DOM in a unified mode, the number of rearrangement and redrawing times of a browser is reduced, CPU and memory occupation is further reduced, and the canvas still keeps smooth response even in the scene with dense elements and frequent updating.

The invention relates to the technical field of data processing, in particular to a cloud platform data analyzing and processing system based on front-end segmentation, which comprises the steps of synchronously displaying fields and paging information to construct export parameters, cloning rendering data to establish a running state, monitoring page turning change to generate buffer content, and segmenting and packaging buffer data according to a memory state. And pushing the data to the local, rolling back the page structure, and completing the export of the process ending identifier. According to the method, export parameters are constructed by synchronously searching fields, header structures and paging numbers; a data clone structure is established in combination with field mapping and paging tracks; field changes are monitored to generate buffer contents; paging data are split according to memory occupation and packaged into local caches; and pushing data to the local and rolling back a display structure to support key links of export processes such as field tracking, memory control, data merging and page recovery.

A data processing apparatus, a processor, a board card, and a data processing method. The data processing apparatus may be comprised by a combined processing apparatus (20), the combined processing apparatus (20) comprising a computing apparatus (201), an interface apparatus (202), a processing apparatus (203) and a storage apparatus (204). The computing apparatus (201) is configured to execute an operation specified by a user, so as to execute deep learning or machine learning calculation. The computing apparatus (201) may interact with the processing apparatus (203) by means of the interface apparatus (202), so as to jointly complete the operation specified by the user. The interface apparatus (202) is used to transmit data and a control instruction between the computing apparatus (201) and the processing apparatus (203). The storage apparatus (204) is used to store data of the computing apparatus (201) and the processing apparatus (203). The data processing apparatus provides a vector computation scheme integrating different types of fine-grained quantization, so that processing can be simplified, and the advantages of a low bit width operation and small memory occupation of a fine-grained quantization format are fully utilized.

The invention provides a long text abstract generation method and device, and belongs to the technical field of natural text processing, the method comprises the following steps: using Euclidean norm to carry out importance sorting on Tokens and carrying out compression storage according to a sparse rate, so that global key information is completely reserved and memory occupation is obviously reduced; in the decoding stage, local attention scores and global attention scores are calculated in parallel, entropy differences are mapped into fusion weights through Sigmoid by combining temperature adjusting parameters, and dynamic balance of local details and long-distance dependence is achieved. The local key value pairs and the global key value pairs are subjected to weighted integration based on the fusion weight, a continuous semantic spectrum is formed in a single decoding layer, splicing breakage caused by traditional partitioning is eliminated, the problems of input limitation and semantic splitting are effectively relieved, the context length capable of being processed by a model is expanded under the condition that the calculation amount is not remarkably increased, and the method has the advantages of being simple in structure and convenient to operate. And local and global context information is adaptively fused, so that the accuracy and continuity of the abstract are effectively improved.

The invention discloses a class preloading method and device for Java micro-service cold start acceleration, and belongs to the technical field of cloud computing and Java virtual machine optimization. The method comprises the steps of class dependency analysis based on a historical call chain, hierarchical class preloading execution and shared memory snapshot management. The device comprises a class analysis engine, a preloading decision maker and a memory snapshot management and JVM optimization adaptation module. Through intelligent analysis of class popularity, hierarchical preloading and combination of a cross-JVM instance memory sharing technology, the Java micro-service cold start time is reduced from average 4.2 s to 0.8 s, memory occupation is reduced by 45%, compatibility with a standard JVM and a mainstream framework is kept, and the method is suitable for financial transaction, real-time recommendation and other micro-service applications in a Serverless scene.

The invention relates to the technical field of large model reasoning, in particular to a reasoning performance optimization method and system based on a large model, and the method comprises the following steps: initializing a reasoning performance optimization agent; collecting hardware environment indexes in real time, wherein the hardware environment indexes comprise a video memory utilization rate, a CPU (Central Processing Unit) exchange number, residual video card resources, storage IOPS (Input / Output Per Second) and network throughput; the method has the beneficial effects that related statistical indexes, including model types, model weight file total volume, model average sequence length, the number of tokens per second output by the model, first token time of the model, a display card list occupied by the model, the size of a KVcache block, the size of a KVcache sliding window and the like, of each model in a service system are comprehensively collected and analyzed; the system performance is comprehensively evaluated, and the defects that in the prior art, performance evaluation is not comprehensive, and a real-time monitoring mechanism for key indexes such as memory occupation and network bandwidth is lacked are overcome.

The invention discloses a high-capacity file encryption method and system based on block processing. According to the method, a high-capacity file is read and encrypted in blocks according to the fixed size, the situation that the high-capacity file is wholly loaded to a memory is avoided, memory occupation is reduced, and efficient encryption is achieved; two symmetric encryption algorithms of SM4 and AES are supported, and the corresponding algorithm is automatically identified and called during decryption by embedding an algorithm identifier in an encrypted file; and generating a check code for the encrypted file by using SM3 for integrity check before decryption. A file path coding mode irrelevant to a platform is adopted, a Unicode path is supported, and mainstream operating systems such as Windows and Linux can be compatible; and meanwhile, a buffer area and a multi-thread mechanism are introduced, so that file I / O operation and encryption calculation are executed in parallel, and the encryption efficiency is improved by utilizing the advantages of a multi-core processor. According to the method, the problems of high memory occupation, inconvenience in algorithm switching, poor cross-platform compatibility and low encryption efficiency of high-capacity file encryption are solved, and the method has a wide application prospect.

The invention relates to the field of data management, and provides an equipment label information hierarchical storage method and device based on a prefix tree structure, the method comprises the following steps: caching equipment label information into a memory in the prefix tree structure, the prefix tree structure being a data structure combining a prefix tree and a linked list, each cached node holds a pointer pointing to a child node storage space on a disk; when the amount of data cached in the memory reaches a preset threshold value, newly adding or modifying the equipment label information cached in the prefix tree into a physical log file; after the newly-added or modified operation is completely written into the physical log file, the equipment label information in the memory is stored in the prefix tree file, and the prefix tree file is stored on a disk by adopting a structure of fusing a prefix tree with a B + tree. According to the invention, the problem of overlarge memory consumption of a full-memory storage scheme in the prior art is solved, and the memory occupation is obviously reduced.

The invention belongs to the technical field of edge computing, and discloses a lightweight AI model adaptive deployment system oriented to edge computing. A dependency intensity matrix is constructed by accurately quantifying a data dependency relationship between model layers, network stability is monitored in real time, an abnormal time period is identified, a communication penalty factor and a bandwidth attenuation factor are calculated based on time delay peak distribution characteristics, and then a communication overhead prediction model for cooperative calculation between edge devices is constructed. The system innovatively generates a multi-granularity model segmentation candidate scheme set, identifies a segmentation boundary triggering resource competition by monitoring and calculating load fluctuation and memory occupancy change, and dynamically adjusts a model partition granularity and a mapping strategy according to a gradient transmission quantity and an activation value size at the segmentation boundary. According to the invention, the overhead of cross-device communication is reduced, the utilization balance of computing resources is improved, and the real-time response capability in a time delay sensitive scene is ensured.

The invention relates to the technical field of heat supply network metering supervision, and discloses a heat supply network metering data supervision system based on edge calculation. The system comprises an edge data acquisition module, a metering task management module, an edge resource scheduling module and a heat supply network supervision execution module. The edge data acquisition module acquires original data from a heat supply network metering sensor in real time, and sends the original data to the metering task management module after verification and format conversion; the latter receives the data, generates and stores a metering task record according to the type and the supervision rule, and updates the state to be to-be-processed; the edge resource scheduling module queries a to-be-processed task, monitors the CPU utilization rate, the memory occupancy rate and the network delay of an edge node, determines an execution sequence in combination with the number of tasks and generates a resource allocation scheme; and the heat supply network supervision execution module activates a corresponding node, obtains a task record, executes data analysis and supervision logic processing, and stores a result. The system optimizes data processing and resource configuration by means of edge computing, and adapts to heat supply network metering data supervision requirements.

The invention provides a scheduling method and device for big language model reasoning and electronic equipment. The method comprises the steps that multiple micro-batches of a big language model executing reasoning task are determined; dividing the parameter weight of each layer in the large language model into a plurality of weight pages according to the micro-batch number; and executing each micro-batch of the reasoning task, pre-fetching a preset number of weight pages required by execution of subsequent micro-batches while executing each micro-batch, and storing the pre-fetched weight pages in a GPU memory. When each micro-batch is executed, a subsequent micro-batch weight page is pre-fetched and stored in a GPU memory, and a micro-batch processing process is performed based on the pre-fetched weight page, so that the parameter weight of the whole large language model is prevented from being integrally loaded to the GPU at one time, the calculation waiting time caused by weight loading is shortened, the memory occupation is reduced, and the calculation efficiency is improved. Unnecessary data transmission and storage are avoided, and the task processing efficiency based on the large language model is improved.

The invention discloses a file pre-reading method and device, a medium and equipment, and the method comprises the steps: firstly obtaining a file type of a target file, and executing pre-reading processing through a pre-reading window only when the target file is a continuous file; and dynamically adjusting the size of the window in combination with the hit condition of the pre-reading content on the basis: increasing the window in hit to improve the sequential reading efficiency, and reducing the window in miss to reduce the invalid pre-reading proportion. Through the above self-adaptive adjustment mechanism based on hit feedback, the size of the pre-reading window can be flexibly optimized for different access modes, and excessive pre-reading caused by a fixed pre-reading window in a random access or access mode change scene is avoided, so that unnecessary memory occupation and input and output resource waste are reduced, and the stability of reading performance is kept.

The invention discloses an online real-time editing method and device supporting million-level vector data and a medium, and relates to the technical field of spatial data co-processing, and the method comprises the steps: carrying out the dynamic rendering of a standardized element sequence based on a real-time state notification, rendering the operable standardized elements into green, and carrying out the dynamic rendering of the standardized element sequence; the standardized elements operated by multiple users are rendered to be yellow, and the standardized elements which cannot be operated are rendered to be red; when the standardized element of the user operation is yellow or red, real-time interception is carried out, a prompt is pushed, and a structured operation log is generated; and verifying the adjacency, connectivity and conflict-free inclusion relationship between the element change coordinates in the structured operation log and the topological snapshot index, generating a processing patch by identifying topological breaking points and executing spatial interpolation connection, marking an element coordinate set needing to be processed, and outputting coordinates of a reprocessing region by calculating a minimum bounding rectangle. According to the method, the memory occupancy essential optimization is realized through a density self-adaptive streaming processing mechanism.

The invention discloses a large model fine tuning optimization method based on multi-strategy fusion, which comprises the following steps: designing a dynamic parameter selection mechanism, adaptively determining a parameter subset needing fine tuning according to a task demand and a model structure, and reducing unnecessary parameter updating calculation; constructing a dynamic low-rank decomposition framework, dynamically adjusting the rank of a low-rank matrix according to a model training state and data characteristics, and keeping key information while compressing a parameter scale; a self-adaptive task sensing mechanism is introduced, a fine adjustment strategy is automatically adjusted according to different task characteristics, and the adaptability of the model to various tasks is improved; and a mixed precision training method is adopted, so that the calculation complexity and the memory occupation are reduced on the premise of ensuring the model precision. According to the method, a parameter efficient fine tuning technology and a dynamic low-rank decomposition strategy are innovatively combined, and an adaptive task perception mechanism and a mixed precision training technology are introduced, so that the operand and resource requirements of model training are effectively reduced, and the fine tuning efficiency and the model performance are improved.

The invention discloses an Internet of Things device anti-quantum password migration method based on a national cryptographic algorithm, and relates to the field of data security, the method comprises the following steps: configuring a Hash agility module taking a national cryptographic SM3 as a default algorithm, and dynamically switching to an anti-quantum or compatibility alternative algorithm based on monitoring; carrying out identity authentication and temporary key negotiation by adopting SM2, and generating a key pair only used for a single session; deriving a session key based on the key pair, and adopting a differential double-layer signature verification mechanism according to the data security level; for firmware updating, a differential package is generated through function level differential analysis, and atomic updating and rollback are achieved through A / B system partition. According to the method, quantum computing threats are effectively resisted, national security compliance and forward security are ensured, the volume of an update package is remarkably reduced through a resource optimization technology, memory occupation is reduced, verification efficiency is improved, and the method is particularly suitable for resource-limited Internet of Things equipment.

The invention provides an AI-based multi-node rapid disaster recovery switching method and system, and the method comprises the steps: obtaining multi-dimensional monitoring indexes of the CPU usage rate, the memory occupancy rate, the network IO, the disk IO, the application response time and the error rate of each node, and constructing a time sequence feature matrix; extracting space correlation features among nodes by using a graph convolutional network, extracting time sequence features through a long-short-term memory network, constructing a deep space-time graph neural network to identify a fault precursor, and predicting a node fault 30-120 seconds in advance; generating a node health score based on a node fault prediction result and capacity estimation, executing a target node optimization algorithm, and selecting an optimal disaster recovery backup node; the method comprises the following steps: constructing a hot backup by adopting a CRI U-based lightweight process migration technology, and generating a check point and an incremental snapshot of a source container; and automatic disaster recovery switching is executed before the node fails, the service flow is redirected, and the service is ensured not to be interrupted. According to the method, the disaster recovery switching time can be shortened to a millisecond level, and the system availability is remarkably improved.

The invention relates to the technical field of computer vision, and discloses an industrial product visual defect detection method and system based on multi-task reverse knowledge distillation, by introducing a multi-scale projection layer and a multi-task learning framework, the expression ability of a model to normal sample features and the inhibition ability to abnormal signals are significantly improved, and the detection accuracy is improved. Therefore, industry-leading defect detection and positioning precision is realized on industrial reference data sets such as MVTec, extremely low reasoning delay and memory occupancy are kept, the reasoning speed is far higher than that of mainstream methods such as PatchCore, an efficient and accurate technical scheme is provided for industrial large-scale real-time defect detection, low delay and low memory consumption are kept, and the defect detection accuracy is greatly improved. The abnormal detection and positioning precision of industrial product visual defect detection is remarkably improved, and the defect that performance and efficiency are difficult to consider in the prior art is overcome.