56 results about "Numerical precision" patented technology

The invention discloses a large language model low-delay reasoning method based on dynamic reasoning graph optimization, and provides a low-delay reasoning method based on dynamic reasoning graph optimization. Constructing a template inference graph which can be rewritten and replayed, and establishing a template library according to input shape vectors; during reasoning, a template is matched with a distance threshold value, and only the attention / feedforward sub-graph is locally recaptured when the distance threshold value exceeds the threshold value; executing a forward execution graph, injecting a key value cache page pointer, numerical value precision and an adapter identifier, and performing playback; dividing the pre-filling and decoding sub-graphs to implement graph-level scheduling; switching the key operator when the key operator operates between a standard / fast kernel and different precisions; page-level backspacing of speculative branches is achieved through a shadow page table and reference counting, and batch and template selection is adjusted in a self-adaptive mode based on online indexes; compared with the prior art, the method has the advantages that the recapture and start overhead is reduced, tail delay and jitter are inhibited, and the hardware utilization rate and the service stability are improved.

The invention provides a joint prediction method and system for a wind resistance coefficient and a surface pressure field of an automobile, and relates to the technical field of automobile aerodynamics. According to the joint prediction method, the pressure field and the wind resistance coefficient are predicted in the same network in parallel, and the wind resistance coefficient obtained by pressure field integration and derived from the pressure field and the wind resistance coefficient obtained by network direct regression are calibrated by physical consistency constraint. The model has physical confidence and interpretability while ensuring numerical precision, so that a key stress area can be positioned and serves an automobile modeling iteration process.

The invention provides a model reasoning method and device, electronic equipment, a storage medium and a program product. The method comprises the following steps: performing attention calculation on inferred tokens by adopting first numerical precision through a large language model to obtain a first attention score corresponding to each inferred token; screening a target tokens from the inferred tokens based on the first attention score; reasoning the input sequence by adopting second numerical precision through the large language model to obtain a reasoning result output by the large language model; the input sequence comprises a target token and tokens to be input corresponding to the token to be inferred; the tokens to be input are tokens pre-selected from the inferred tokens according to a preset rule; the first numerical precision is lower than the second numerical precision. According to the method, the sparse processing of the token is realized by adopting mixed precision calculation, so that the reasoning efficiency of the large language model is improved.

The invention discloses a split type three-box girder nonlinear vortex vibration steady-state amplitude intelligent prediction method, belongs to the field of bridge vortex-induced vibration control, and aims to solve the problem of low accuracy of three-box girder vortex vibration nonlinear modeling and amplitude prediction. The method comprises the following steps: acquiring displacement time sequence data through a reduced scale model wind tunnel test, preprocessing to obtain displacement, speed and acceleration dimensionless data, and dividing a data set; constructing a candidate function library containing high-order polynomial terms of the primary function; performing nonlinear system sparse recognition by taking acceleration data as a target item and combining an improved algorithm to obtain a preliminary feature set; a final feature set is obtained through energy-statistics-self-adaption three-stage screening; and finally, constructing a control equation to predict the steady-state amplitude. The method can accurately predict amplitude, coincide test and real bridge observation results, and is less in computing resource occupation and high in numerical value precision.

The invention discloses a photon counting CT image noise reduction method based on multi-channel WGAN, and relates to the technical field of medical image processing, and the method comprises the steps: taking high-energy, low-energy and all-round CT images as generator input, extracting multi-scale features through independent convolution branch, carrying out the global average pooling of each channel feature map, and calculating the attention weight between channels to carry out the weighted fusion; a WGAN structure is adopted, and a discriminator is constrained through a Wasserstein distance and a gradient penalty term so as to improve the training stability; meanwhile, the mean square error, the structural similarity index and the perception loss are integrated, and the numerical precision and the visual effect of the noise reduction image are balanced; and by alternately optimizing the generator and the discriminator, performing multi-channel feature processing and up-sampling on the noisy low-dose CT image after training is completed, and outputting a high-quality noise-reduced image.

The invention relates to the technical field of data storage, and discloses a numerical data variable-length coding storage method, device, equipment and medium, which comprises the following steps: converting target numerical data into fixed-length binary representation according to definition precision, executing filling removal interception processing on an integer part, executing digital-level packet compression coding on a decimal part, and performing variable-length coding on the target numerical data; and generating a variable-length binary sequence containing length information and symbol information, and writing the sequence into a storage medium to realize compact storage of numerical data. According to the method, redundant filling bytes are eliminated from an integer part, compressed encoding is performed on a decimal part according to effective digits, and unified processing of sign bits is combined, so that a storage space dynamically changes along with actual numerical digits on the premise of keeping numerical precision and analyzability, space waste of fixed-length numerical types is reduced, and the method is suitable for large-scale popularization and application. Meanwhile, the precision error of the floating-point number is avoided, and the storage efficiency and the data reliability are improved.

The invention relates to an icing power transmission line galloping simulation method and simulation system based on a composite time integral. The method comprises the following steps: S1, establishing a nonlinear finite element model of an icing power transmission line; s2, carrying out time domain solution by adopting a composite time integration algorithm; calculating the solution of the moment in two sub-steps according to the solution of the moment, namely from the first sub-step to the second sub-step; the first sub-step adopts a trapezoidal rule for integration, and the second sub-step adopts a three-point backward difference format based on information of three moments; s3, coupling and solving a nonlinear aerodynamic load related to the speed; s4, obtaining the galloping response of the power transmission line; the problems that in a transmission line galloping simulation technology based on a traditional Newmark method, contradiction between calculation efficiency and numerical precision exists in the strong nonlinear coupling problem, accurate tracking of low-frequency galloping and efficient suppression of high-frequency noise are difficult to achieve at the same time under the large time step length, simulation calculation efficiency is low, and convergence is poor are solved.

The invention relates to the technical field of artificial intelligence, in particular to a cache management method, device, equipment, medium and product, and the method comprises the steps: evaluating the importance scores of different attention heads for a current reasoning task in real time, and dynamically allocating different levels of storage resources for the key value cache of each attention head according to the importance scores, comprising a storage position and numerical precision, so that the memory is saved to the greatest extent on the premise of ensuring the model output quality and particularly not losing key information, and the reasoning speed is further improved.

The invention relates to a transient electromagnetic three-dimensional inversion method and device, computer equipment and a storage medium. The method comprises the following steps: decoupling a forward modeling grid and an inversion grid, independently generating a vertical self-adaptive layered special forward modeling grid for each time channel according to the skin depth of the time channel, and remarkably reducing the calculation dimension while keeping the horizontal discrete consistency; an equivalent circuit method is innovatively adopted to achieve high-precision mapping of model parameters from an inversion space to a forward modeling space, a plurality of inversion grid units arranged in the current magnetic field component direction are regarded as series connection, and a plurality of inversion grid units arranged in the direction perpendicular to the current magnetic field component direction are regarded as parallel connection. The equivalent conductivity in the control volume is accurately calculated by using a series-parallel resistance formula, and the precision defect of a traditional method in a complex electrical structure is effectively solved. According to the method, the calculation efficiency and the numerical precision are effectively balanced, and the optimal discretization modeling of the transient electromagnetic response is realized.

The invention provides a task execution method which comprises the steps that according to a first query feature and a first key feature, a computing unit is used for executing a first multiplication task of a target attention task queue, a first multiplication result feature is obtained, the first key feature is obtained by converting a second key feature of a fixed point type into a floating point type, and the second key feature is obtained by converting a second key feature of a floating point type; the numerical precision of the first key feature is greater than that of the second key feature, and the numerical precision of the first query feature is less than that of the second query feature; according to the first multiplication result feature and the first value feature, executing a plurality of target tasks of the target attention task queue by using a computing unit to obtain attention features; the first query feature is obtained through the second query feature and an inverse quantization scaling parameter for the second key feature, so that the hardware resource overhead of executing the target attention task queue by the computing unit is smaller than a preset hardware resource overhead value. The invention further provides a task execution device and equipment, a medium and a program product.

The invention provides a multidisciplinary coupling numerical simulation generality algorithm library and a parallel computing method, and relates to the field of aviation multidisciplinary coupling numerical simulation. Comprising a parallel coupling format library, a data mapping library, a grid support library, a data bus library, an MPI parallel library and a subject calculation interface library which are used for providing common algorithm support in multidisciplinary coupling simulation. According to the parallel computing method, a two-layer parallel architecture based on multiple plug-ins, multiple data and MPI is adopted, parallel in subjects, parallel between subjects, distributed data exchange and cross-process communication are achieved, and the parallel efficiency and numerical precision of coupling simulation are improved. The method is suitable for pneumatic + X type multidisciplinary coupling simulation, supports flexible expansion of disciplinary types and number, and has good parallel expandability and coupling precision.

The invention relates to an interpolation method and an interpolation model for scRNA-seq data. An interpolation method for scRNA-seq data comprises the following steps: performing dropout interpolation on the scRNA-seq data through an automatic encoder of a Markov self-attention mechanism and a self-adaptive threshold low-rank approximation method to obtain an interpolation matrix XCG and an interpolation matrix XMF; the interpolation matrix XCG and the interpolation matrix XMF are integrated through a cell-level interpolation matrix integration method, and an interpolation result is obtained; and taking the interpolation result as the input of the next iteration until the iteration condition is met, and obtaining the final interpolation result after the loop is ended. According to the interpolation method and the interpolation model for the scRNA-seq data, the global heterogeneity of the scRNA-seq data is ensured, meanwhile, the local structure of the scRNA-seq data is captured, the interpolation value is highly consistent with the real data in trend, and the optimal numerical precision is also obtained.

The invention discloses a discontinuous displacement fracture phase field calculation method for a composite material. The method comprises the following steps: S1, establishing a strain energy density model of the composite material; s2, based on the strain energy density model, calculating a displacement control equation, a phase field control equation and a damage control equation according to a variational method; s3, performing one-dimensional analysis on the strain energy density model, the stress in the displacement control equation, the phase field control equation and the damage control equation to obtain an analytic equation; and S4, calculating the discontinuous displacement fracture phase field of the composite material based on the analytical equation. And S5, compiling the steps S1 to S4 into a UEL user subprogram, integrating the UEL user subprogram to an ABAQUS finite element platform, and outputting a crack path diagram of the composite material. According to the method, fracture behaviors such as interface debonding facing the composite material can be correctly captured without a particularly small grid scale while the numerical precision is ensured, and a composite material crack path prediction map is obtained.

In order to solve the problems that in the prior art, calculation precision and efficiency are difficult to consider at the same time during supersonic jet machining, and detached shock waves cannot be accurately captured, the invention provides a finite element simulation method for the stagnation flexible constraint flow field characteristics of a jet impact area. The finite element simulation method comprises the steps that firstly, a geometric model of a jet area and an outer watershed is established, and preliminary solving is conducted; extracting jet flow axis pressure gradient characteristics to determine detached shock wave interval coordinates; then reconstructing a block geometric model comprising an upper jet flow region, a detached shock wave region, a stagnation region and an outer watershed according to the coordinates, and implementing differential local grid encryption on the shock wave region and the stagnation region; and finally, carrying out second iterative solution based on the Sutherland viscosity law to obtain final flow field distribution. According to the method, the thickness and the position of the shock wave layer are accurately captured while the number of the grids is controlled through the block encryption strategy, the contradiction between the calculation economy and the numerical precision under the transonic flow condition is solved, and the method is suitable for numerical simulation of the complex supersonic gas jet flow field containing the detached shock wave and the stagnation region.

The invention discloses a gradient numerical calculation method and system based on grid self-adaption and a hybrid strategy, and the method comprises the steps: obtaining all grid information in a to-be-calculated region, and carrying out the recognition of grid positions; judging whether the grid belongs to an internal grid or not according to the identified grid position; if the grid does not belong to the internal grid, grid center gradient calculation is carried out based on a Green-Gauss method, and a grid center gradient is obtained; if the grid unit belongs to the internal grid, calculating the grid distortion of all the internal grids based on a grid distortion calculation method of different types of grid units; calculating a weighting coefficient according to the grid torsion degree; according to different weighting coefficients, calculating grid core gradients by adopting different grid core gradient calculation methods; and calculating a gradient value of the interface based on the grid core gradient of the adjacent grid units of the interface. According to the method, the numerical precision, the adaptability, the stability and the calculation efficiency of the overall grid gradient are improved, and a new thought is provided for high-precision numerical discrete solution of computational fluid mechanics.

The application provides a combined prediction method and system for a wind resistance coefficient and a surface pressure field, and relates to the technical field of automobile aerodynamics. The combined prediction method predicts the pressure field and the wind resistance coefficient in parallel in the same network, and calibrates the wind resistance coefficient derived from the pressure field obtained by pressure field integration and the wind resistance coefficient directly regressed by the network with physical consistency constraints, so that the model has physical credibility and interpretability while ensuring numerical accuracy, thereby being capable of locating a key stress area and serving an automobile modeling iteration process.

The application discloses a kind of double box girder flow field dynamics modal and flow stability analysis method and system, belong to bridge engineering field.This method first establishes double box girder flow field model by computational fluid dynamics, solves the momentum equation and continuity equation of two-dimensional incompressible flow;Dynamic modal decomposition is carried out again, and eigenvalue, characteristic mode and other information are obtained;Finally, based on linear stability theory, the flow field component is expressed as the sum of steady-state solution and disturbance term, and the global linear flow stability equation is obtained by substituting into the equation, and the flow stability is analyzed by discrete.This method can accurately extract the modal characteristics of each order, providing a scientific basis for wind-induced vibration analysis of double box girder, and has important application value.The method of the application occupies less computing resources, has high numerical accuracy, and can be further expanded to higher precision calculation model according to computing resources, and is suitable for application in actual bridge engineering.

The application relates to the technical field of engineering simulation, and discloses an industrial structure mechanics simulation prediction system based on an enhanced graph attention network, which comprises a finite element data mapping and enhancement module, a prediction module, an optimization module and a deployment module. The system reconstructs a finite element grid into a graph structure tensor, captures geometric mutation characteristics by using high-dimensional feature projection and adaptive feature recalibration, and transmits deep physical information through dense residual connection. In the training stage, a physical constraint loss function containing attention area weighting and structure smoothness is introduced; in the reasoning stage, a mixed precision mode is adaptively switched according to hardware attributes, and large graph blocking and fusion calculation based on overlapping boundaries are performed. The application effectively solves the contradiction between calculation efficiency and numerical accuracy in complex industrial structure mechanics response prediction, and realizes millisecond-level high-fidelity online simulation.

The application discloses a large language model low-delay inference method based on dynamic inference graph optimization, and proposes a low-delay inference method based on dynamic inference graph optimization; a rewritable and replayable template inference graph is constructed, and a template library is established according to an input shape vector; during inference, a template is matched according to a distance threshold value, and only attention / forward subgraphs are locally recaptured when the threshold value is exceeded; a forward execution graph injects a key-value cache page pointer, a numerical precision and an adapter identifier, and is replayed; pre-populated and decoded subgraphs are divided to implement a graph-level scheduling; a key operator is switched between a standard / quick kernel and different precisions during operation; a page-level rollback of a speculative branch is realized through a shadow page table and a reference count, and a batch and a template selection are adaptively adjusted based on online indicators; compared with existing schemes, the application reduces recapture and startup overhead, suppresses tail delay and jitter, and improves hardware utilization and service stability.

The application provides a steel structure calculation method and system based on physical logic anchoring and heterogeneous decoupling, and relates to the technical field of intelligent steel structure calculation. The method comprises the following steps: extracting user's unstructured engineering requirements into a structured feature vector, constructing a directed acyclic computation graph along a knowledge graph; starting knowledge graph reasoning and retrieval enhancement in parallel to generate a two-way evidence, calculating a logic anchoring coefficient Lambda, which includes a number version consistency judgment, a specification timeliness dynamic weight and a strong item penalty, comparing Lambda with a dynamic reliability threshold and executing a three-branch decision to eliminate index illusion; routing the calculation instruction to a physically isolated sandbox after triple audit to determine the solution and eliminate calculation illusion; based on the logic dependence hypergraph, the dissenting node is traced back and locally recalculated, and a compliant calculation report is output. From the architecture level, the calculation illusion and index illusion in the steel structure calculation are eliminated, and the industrial numerical precision and reliable specification compliance traceability are realized.

The present application relates to the technical field of computers. Provided are a floating-point number processing method, and a computing device and a computing chip. The computing device acquires a first floating-point number, and obtains a second floating-point number on the basis of a first sign field, a first exponent field and a first mantissa field of the first floating-point number. The data format used by the second floating-point number comprises a shared data part and a variable data part, wherein the shared data part is stored in a shared area of a memory, and the shared data part is shared by a plurality of pieces of data within the data range to which the first floating-point number belongs. In the shared area of the memory, a common shared data part may be stored for a plurality of data, thereby facilitating the representation of floating point numbers by using limited bit widths. The shared data part comprises a second exponent field and a second mantissa field, wherein the second mantissa field of the shared data part may be used for compensating for the precision of values that can be represented by the variable data part, thereby improving the numerical precision of data that can be represented by the floating-point numbers, and thus satisfying the requirements of AI training and inference.

Techniques for computing matrix operations for arbitrarily large matrices on a limited size hybrid analog-digital matrix processor are described. Techniques for gain adjustment in a limited size hybrid analog-digital matrix processor are described that enable the system to achieve higher energy efficiency, greater physical density, and improved numerical accuracy. In some embodiments, these techniques maximize the prediction accuracy of GEMM-based convolutional neural networks using low-precision data representations.

The invention relates to the technical field of a cesium feed-in system in a magnetic confinement nuclear fusion negative neutral beam system, in particular to an optimization method for DSMC numerical simulation of a cesium feed-in system based on CUDA parallel. According to the technical scheme, the method comprises the following steps: setting a molecular initial state at a host end, and preparing initial conditions for analog computation; allocating a memory space for the molecular array and the related data structure at the equipment end; molecular motion is calculated in parallel through a GPU at an equipment end, and collision between molecules and a wall surface is processed. According to the method, the simulation efficiency is improved to the minute level through a CUDA parallel computing architecture, a grid-based molecular index and memory access mechanism is optimized, the statistical reliability is guaranteed by adopting a thread-independent random number generator, near-vacuum molecular index is accurately restored by means of a decoupled physical model, and the statistical reliability is improved. Therefore, an efficient and practical analysis tool is provided for cesium feed-in system engineering design on the premise that the numerical precision is guaranteed.

The invention relates to a rapid target detection method based on an RK3588 platform, and solves the problem that the real-time requirement is difficult to meet because the frame rate is often lower than 10fps when a target detection algorithm is directly operated in equipment. According to the method, the operation efficiency of a target detection algorithm on an RK3588 platform can be greatly improved through the methods of modifying a model structure and an activation function, quantifying the model, using a thread pool and the like on the basis of a yov8s target detection model, and the actual measurement frame rate can exceed 100fps. And finally, high-frame-rate target detection on the edge computing equipment is realized. The method specifically comprises; the method comprises the following steps: 1) modifying a model structure, replacing an activation function, removing a decoding operation on a bounding box in a YOLOv8 model, and executing the decoding operation on the bounding box in a CPU (Central Processing Unit); 2) quantizing the model trained by the RK3588 platform so as to better exert the performance of the NPU, specifically including (1) numerical precision reduction, (2) linear transformation, (3) thread pool acceleration: accelerating model reasoning by using a thread pool, and (4) target detection.

The invention discloses an unconditionally stable UPML electromagnetic simulation method based on Dosin FDTD, and belongs to the field of computational electromagnetics and numerical analysis, and the method comprises the steps: introducing an anisotropic damping term under a Dosin finite difference time domain (MS-FDTD) framework, embedding the UPML absorption characteristic into a Dosin Maxwell system, and carrying out the UPML electromagnetic simulation. And the consistency of the main domain and the absorption layer in geometric structure and energy transfer is realized. By adopting a Preissmann implicit discrete format and a local one-dimensional operator splitting strategy, a numerical value updating form irrelevant to a time step and a space step is obtained, so that unconditional stable solution is realized. According to the method, the electromagnetic field quantity is decoupled through the intermediate variable of the electric displacement vector and the magnetic induction intensity, and the solving efficiency and the numerical precision are improved. The provided Dosin UPML method can keep energy conservation and structural stability in complex media and long-time simulation, and the electromagnetic wave absorption precision and simulation reliability are remarkably improved.

The embodiment of the invention provides an operator quantification method and device based on mixing precision, equipment and a storage medium, and belongs to the field of model quantification. The method comprises the following steps: acquiring an operator feature data set of a neural network model to be quantized, wherein the operator feature data set comprises at least two quantitative indexes of the sensitivity of each operator in the neural network model to quantization; based on a preset clustering algorithm, performing clustering processing on the operator feature data set to obtain at least two clustering clusters, and classifying operators corresponding to feature data in each clustering cluster into one group to obtain at least two operator groups; according to at least two quantitative indexes of the clustering center of each clustering cluster, distributing corresponding numerical value precision to each operator in each operator group; and according to the numerical value precision corresponding to each operator, performing quantification processing on the neural network model. According to the method, appropriate numerical value precision can be accurately distributed to different operators in the neural network model, so that the precision of the quantized neural network model is ensured.

The invention relates to the technical field of electric power simulation, and discloses a data splitting processing method for an electromagnetic simulation model of an electric power system. The method comprises the following steps: constructing a node admittance matrix and forming an electrical distance matrix through an eigenvalue decomposition algorithm; based on the electrical distance matrix, executing a K-means clustering algorithm to divide the node into a plurality of electrical coupling subsystems; identifying the boundary connection relationship between the subsystems to determine boundary nodes and boundary branches; performing Thevenin equivalence processing on the boundary nodes and the boundary branches, and splitting a node admittance matrix to generate independent admittance sub-matrixes and equivalence boundary parameters; and distributing the independent admittance submatrixes and the equivalent boundary parameters to different calculation units to realize parallel electromagnetic transient simulation. The problems of low parallel computing efficiency and insufficient boundary processing precision caused by an unreasonable data splitting strategy in electromagnetic transient simulation of an existing power system are solved, and the computing efficiency and the numerical precision of electromagnetic transient simulation of a large-scale power system are improved.

The invention discloses a method and a system for analyzing dynamic modality and flow stability of a flow field around a double-box girder, and belongs to the field of bridge engineering. The method comprises the following steps: firstly, establishing a double-box girder surrounding flow field model through computational fluid dynamics, and solving a momentum equation and a continuity equation of a two-dimensional incompressible flow; performing dynamic mode decomposition to obtain information such as characteristic values and characteristic modes; and finally, based on a linear stability theory, representing a flow field component as a sum of a steady-state solution and a disturbance term, substituting the sum into the equation to obtain a global linear flow stability equation, and analyzing the flow stability through discretization. The method can accurately extract modal features of each order, provides a scientific basis for double-box girder wind-induced vibration analysis, and has important application value. The method occupies few computing resources, is high in numerical precision, can be further expanded to a computing model with higher precision according to the computing resources, and is suitable for being applied to actual bridge engineering.

The invention provides a tensor calculation coprocessor which comprises a preprocessing remapping module, a scheduling calculation module and an output mapping module. The preprocessing remapping module is used for receiving task parameters of a tensor calculation task transmitted based on a preset bus and matrix data of each to-be-processed matrix and mapping the task parameters and the matrix data to the input cache unit; the scheduling calculation module is used for scheduling the task parameters from the input cache unit to the target calculation array and scheduling the matrix data to the target calculation array in a blocking manner for tensor calculation to obtain a calculation result and a block index; and the output mapping module is used for mapping the calculation result to the output cache unit according to the block index. The coprocessor supports hardware data preprocessing remapping, an AXI universal interface and flexible parameter configuration of the coprocessor and supports various numerical value precisions and matrix calculation sizes, the overall calculation efficiency of the tensor calculation coprocessor is improved, and the comprehensive advantages of high flexibility and high adaptability are achieved.

The invention relates to a target identification method and system based on edge computing, and the method comprises the steps: deploying a model framework and an ecological system component on an edge device, obtaining a trained OfficientDet-Lite model, carrying out the format conversion, quantization and hardware acceleration optimization, and deploying the model framework and the ecological system component in the edge device; the quantization comprises performing low-precision conversion on the model parameters of the model and the numerical precision output by the middle layer; the hardware acceleration optimization comprises unloading all operations of the model to a hardware accelerator on the edge equipment for execution; and receiving the to-be-identified data information through the edge device, inputting the to-be-identified data information into the deployed model, and obtaining a target identification result. Compared with the prior art, the network model optimization is carried out aiming at the resource limitation of the edge equipment, and excellent balance is provided between the precision and the efficiency.