853 results about "Numeric Value" patented technology

The invention provides a language model training method and device, a text processing method and device, equipment and a medium, and relates to the technical field of natural language processing, and the method comprises the steps: predicting a first probability matrix corresponding to each data unit in a sample text based on a teacher model; the first probability matrix comprises a probability value of each data unit belonging to each lexical element in the first word list; according to the numerical value of each probability value in the first probability matrix, compressing the first probability matrix to obtain a second probability matrix corresponding to each data unit; performing alignment operation on the second word list according to lexical elements corresponding to probability values in the second probability matrix to obtain a third word list; and according to the third word list and the second probability matrix, carrying out distillation training on the student model to obtain a target language model, thereby reducing the storage cost, improving the distillation training efficiency, and enabling the target language model trained according to the method to better adapt to different model architectures and text processing scenes while keeping high performance.

A model compression method includes obtaining a first weight of each layer of a neural network model, where the first weight of each layer is a value of a floating-point type; and quantizing the first weight of each layer based on a quantization parameter to obtain a second weight of each layer. The second weight of each layer is a multi-bit integer. Quantities of bits of second weights of at least a part of layers are different. A quantity of quantization bits of a second weight of a layer with high sensitivity to a quantization error is greater than a quantity of quantization bits of a second weight of a layer with low sensitivity to the quantization error.

The invention discloses a 4HPU-based natural language character and numerical value general solution system and method and application thereof, and aims to solve the technical problem of realizing unification, universality and reasonability of natural language character and numerical value codes. A main coding path for realizing coding generation and structure nesting is arranged; the method comprises the following steps of: constructing a basic coding unit, judging the types and codes of numerical values, letters, Chinese characters or symbols; compared with the prior art, the method has the advantages that natural numbers, letters, Chinese characters and symbols are subjected to structured coding in a right-to-left nesting mode, complete coded word strings are obtained, structural expression and path reasoning are clear, dynamic expansion and compression can be achieved, and the coding efficiency is improved. And a more excellent basic coding method is provided for application scenes of data recovery and reverse recognition, natural language processing, intelligent communication and low-power-consumption equipment.

The invention relates to a dynamic speculation sampling method and system for large model reasoning. The method comprises the following steps: processing requests for multiple groups of input test data of similar scenes based on a predetermined application scene reasoned by the large model, obtaining the maximum throughput data volume of the throughput data volume of the target model in the conventional sampling mode and the maximum throughput data volume of the throughput data volume of the target model in the speculation sampling mode, wherein the maximum throughput data volume is larger than or equal to the maximum throughput data volume of the throughput data volume of the corresponding conventional sampling mode. Configuring a request number span sequence which is used for dynamic speculation sampling and is arranged from large to small and a draft token number sequence which is arranged from small to large and has corresponding length; aiming at a specific data processing request, determining a span value interval formed by adjacent request values of the current survival request number in the request number span sequence according to the current survival request number, and setting a draft token number for each request in the current survival request number based on a draft token value in a draft token number sequence corresponding to a lower request value of the span value interval.

The invention relates to the technical field of data processing, in particular to an LLM-based multi-source heterogeneous data intelligent fusion and self-adaptive processing method, which comprises the following steps of: receiving multi-source heterogeneous data, matching character sequences of mailbox fields and identity card number fields in records one by one by adopting a preset regular expression model, according to a preset key value pair mapping dictionary, items are uniformly converted, data records which cannot pass format verification, type conversion and value domain mapping are marked as doubt, and data records which cannot be processed or are marked as suspicious are separated. According to the method, a hierarchical processing flow is constructed, and deterministic format verification, type conversion and value domain mapping are arranged at the front end of a processing link, so that most simple data quality problems can be quickly processed with low calculation overhead. Then, numeric statistics and character string comparison are used to carry out fuzzy correction on the data, and spelling errors and numeric abnormities which cannot be covered by rules are further solved.

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 part monitoring method, an electronic device, a storage medium and a product, and relates to the technical field of servers, the method is applied to a data processing unit card, the data processing unit card comprises N parts, N is a positive integer greater than or equal to 2, and the method comprises the following steps: obtaining the health state of each part in the N parts; a binary number corresponding to each component is generated based on the health state of each component, N binary numbers are obtained, each binary number is a first numerical value or a second numerical value, the first numerical value is used for indicating that the health state of each component is healthy, and the second numerical value is used for indicating that the health state of each component is abnormal; forming a binary character string by the N binary numbers according to a preset rule; and when a monitoring request sent by a substrate management controller at a host side is received, sending the binary character string to the substrate management controller. Multiple query command monitoring components do not need to be sent frequently, and occupation of a large number of system resources and network bandwidth is avoided.

The embodiment of the invention provides a data processing method for a chip, the chip and electronic device.The method comprises the steps that in response to a target triggering condition, counting values of all performance counters in the chip are grouped and collected, and multiple sets of counting data are obtained; each group of counting data comprises counting values of a plurality of performance counters; performing compression processing on the multiple groups of counting data to obtain multiple groups of compressed counting data; and respectively sending the plurality of groups of compressed counting data to a memory. Therefore, on one hand, the dumping efficiency of the count value of the performance counter can be improved, and the overall performance of the chip is improved; on the other hand, the problem that the counting data of each group cannot be matched easily due to too many groups can be improved, and the accuracy after the counting value data dumped in groups are recombined is improved; and on the other hand, time consumed by the user to wait for numerical value dump of the performance counter can be reduced, and user experience is improved.

A system and method for data compaction and encryption of anonymized data records. A dataset may be pre-processed by dividing into sourceblocks at reasonable intervals and tallying each sourceblock's frequency, creating a tally record of tokens and count values. This tally record may then be anonymized and transmitted to a data deconstruction engine which combined with a library manager creates a codebook and performs optimization techniques on the codebook. The data deconstruction engine and library manager may be distributed across multiple nodes or devices. The received anonymized tally record may be parsed into individual tokens by identifying the tokens with the highest count value. The tokens may then be sent descending order of count value to the library manger where each token may be assigned a codeword. A half-backed codebook is then created using the tokens and each token's unique codeword, before sending the half-backed codebook to a system user.

According to the storage and calculation integrated processor based on the POSIT format and the processing method, the calculation complexity during network training can be greatly reduced by utilizing the improved POSIT data format, meanwhile, due to the fact that the characteristic of dynamic bit width is reserved, the numerical value representation range is widened, the neural network training precision is guaranteed, and the processing efficiency is improved. The problems that in the prior art, low-precision training supports multiple floating points or integer data formats, and POSIT data types are not supported are solved. And a large number of optimization means are needed to ensure the precision, so that the technical scheme is complicated, the data format lacks flexibility, and network parameters cannot be effectively represented.

The invention discloses a method for disguising and hiding data in a memory address, and relates to the technical field of data protection and memory management. The method comprises the steps of obtaining a real numerical value of a target variable needing to be protected; splitting the real numerical value into a plurality of sub-numerical values, and storing the plurality of sub-numerical values in a group of newly allocated storage units in a memory; and when the target variable needs to be used, restoring the real numerical value of the target variable from the plurality of sub-numerical values. The target variable still allocates an address in the memory, but stores the disguise value. The sub-values can be generated through multiple attenuation modes and written into a protection array which is the same as the target variable data in type. The sub-values can be stored in a random index or sequential storage mode, and the protection array can be re-declared during each assignment or reused when conditions are met. According to the method, the hiding performance and the tamper-resistant capability of the data in the memory are effectively improved, and the method is suitable for an application scene in which safety protection is carried out on the local operation data.

The invention belongs to the field of computers, and discloses a variable-precision parallelization index floating point number storage and calculation integrated method based on an FPGA (Field Programmable Gate Array), which specifically comprises the following steps of: inputting decimal floating point data; an index floating-point number binary sequence format is defined, decimal floating-point data values are converted into the index floating-point number binary sequence format, parameters of the index floating-point number binary sequence format comprise sign bits, index bits, mantissa bits, specified index offset and mantissa bit width, and the mantissa bit width is dynamically adjusted according to calculation requirements; based on decimal floating point data numerical values, extracting parameters of an index floating point number binary sequence, and storing the parameters in a register of the FPGA as operands; performing displacement sorting on the two operands stored in the register, and calculating an absolute value of a difference between exponent bits and a difference between mantissa bits of the two operands; the problems that in the prior art, a floating-point number storage scheme is low in calculation efficiency and insufficient in precision in an application scene where the dynamic range changes drastically are solved.

To reduce power consumption, data bits or a portion of a data register that is not expected to toggle frequently can be grouped together, and be clock-gated independently from the rest of the data register. The grouping of the data bits can be determined based on the data types of the workload being operated on. For a data register configured to store a numeric value that supports multiple data types, the portion of the data register being clock-gated may store a group of data bits that are unused for one or more data types of the multiple data types supported by the data register. The portion of the data register being clock-gated can also be a group of data bits that remain unchanged or have a constant value for numeric values within a certain numeric range that is frequently operated on.

The invention discloses an in-memory calculation-oriented model parameter encryption method and device and a model parameter calculation method and device, and relates to the field of data security. The encryption method comprises the following steps: encrypting symbols in model parameters by adopting a symbol key; determining an auxiliary bit line voltage according to the symbol key and a voltage applied to the row line of the storage array; encrypting numerical values in the model parameters by adopting a numerical key; and determining a word line voltage and a bit line voltage according to the symbol ciphertext, the auxiliary bit line voltage, the numerical ciphertext and parameters of the storage array, and storing the symbol ciphertext and the numerical ciphertext in the storage array. The calculation method comprises the following steps: coding a to-be-calculated vector into a bit line voltage and a word line voltage according to a symbol key and a numerical key, transmitting the bit line voltage and the word line voltage to a storage array, decrypting a symbol ciphertext and a numerical ciphertext, and calculating the to-be-calculated vector code to obtain a model result corresponding to the to-be-calculated vector. The safety of the model parameters is improved, and the leakage risk of the model parameters is reduced.

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 discloses a data synchronous scheduling method, a data synchronous scheduling system and a processor. According to the scheme, target data to be written into an FPGA and a data writing address are obtained; data interaction is carried out between the FPGA and the application program through a JTAG interface; updating corresponding data in the intermediate data file based on the target data and the data write-in address, and increasing a numerical value of a counter in the intermediate data file to obtain an updated intermediate data file; the updated numerical value of the counter in the intermediate data file is used for indicating the FPGA to execute data synchronous scheduling with the application program; and based on the updated intermediate data file, executing data synchronous scheduling with the FPGA. According to the technical scheme, the technical problem that effective data synchronous scheduling is difficult to achieve due to the fact that a JTAG interface is limited when data interaction is conducted between an existing FPGA and an application program through the JTAG interface is solved, and the reliability of data synchronous scheduling is improved.

The invention provides a data processing method and device and a computer program product. In the embodiment, to-be-processed data are obtained, numerical values in the to-be-processed data are represented through a first data format, and the first data format comprises a first symbol field, a first order code field and a first mantissa field; according to the precision characteristics of the to-be-processed data and a pre-configured second data format, the order offset of the to-be-processed data is determined, the second data format comprises a second symbol field, an order bit width field, a second order field and a second mantissa field, the order bit width field is used for indicating the bit width occupied by the order field in the second data format in the total bit width of the numerical value, and the second mantissa field is used for indicating the bit width occupied by the order field in the second data format in the total bit width of the numerical value; the precision characteristic of the second data format refers to the precision of numerical values when a second-order code domain of the second data format is in different value ranges; and according to the order offset of the to-be-processed data, converting the data format of the numerical value in the to-be-processed data into a second data format, wherein the bit occupied by the second data format is smaller than the bit occupied by the first data format. Therefore, the scaling operation can be reduced, so that the resource usage amount of a CPU and a memory is reduced.

A computerized method for detection of format drift and format anomalies is described. A format representation for each data point of a first data sample is extracted. Transformations of each format representation is conducted, resulting in a first plurality of count values (reference) and a second plurality of count values. Each count value identifies a number of occurrences of a transformed format representation within that data sample. Thereafter, a first probability distribution for the first plurality of count values and a second probability distribution for the second plurality of count values are computed. Analytics using the first and probability distributions are conducted to produce a first metric. A format drift is determined based on an evaluation of the first metric to a second metric operating as a threshold metric. Format anomalies are detected based on analytics of hashed format representation and determination of infrequent usage of a particular format representation.

A method and apparatus for the identification of suspect behaviour in a retail environment, the method comprising: detecting a person in a frame of said stream of video data; extracting a set of activities of the identified person from the stream of video data; assigning a numeric value to each extracted activity in the set of extracted activities, said numeric value representative of a threat level of the activity; accumulating said numeric values to provide a behaviour score; and identifying a behaviour as being suspect when the behaviour score reaches a target threshold value associated with the behaviour.

The embodiment of the invention provides a distributed data deduplication method and product. The method comprises the steps of obtaining a data key of data to be subjected to deduplication; distributing the data to be subjected to duplicate removal to the ith sub-bucket according to the mixed hash value of the data key; mapping the mixed hash value according to a hash function attribute value configured for the ith sub-bucket to obtain k position numbers; reading numerical values of positions corresponding to the k position numbers from a bit array which is stored at the current moment and corresponds to the ith sub-bucket to obtain k numbers; if it is confirmed that the k numbers are all first numerical values, it is confirmed that the data to be subjected to duplicate removal belong to duplicate data; and if at least one of the k digits is confirmed to be a second numerical value, confirming that the data to be subjected to duplicate removal belongs to new data. According to the embodiment of the invention, the speed and accuracy of big data deduplication processing are remarkably improved, and the problem of single-node calculation overload (different buckets can be distributed to different calculation nodes) is effectively solved.

The invention discloses a dynamic adjustment method of a metadata block, electronic equipment and a storage medium, and relates to the technical field of storage equipment. According to the relation between the frequency proportion of a first erasing frequency and a second erasing frequency and a target preset proportion threshold value determined based on the average erasing frequency and the maximum erasing frequency proportion, the first erasing frequency and the second erasing frequency are adjusted; according to the method, whether the data blocks are adjusted or not is judged according to the first data blocks and the numerical value of the first erasing times, when adjustment is determined, the preset number of data blocks in the first data blocks are adjusted into the new second data blocks, meanwhile, the original second data blocks are merged into the first data blocks, dynamic adjustment of the two types of data blocks is achieved, the erasing times of the two types of data blocks are balanced, and therefore the user experience is improved. The technical problem that the solid state disk enters a read-only state due to the fact that the flash memory blocks used by four stages of units reach the service life limit firstly due to the fact that the difference of the erasing times of the two types of flash memory blocks is too large in the prior art can be solved, and the technical effects of improving the availability and reliability of the solid state disk and prolonging the service life of the solid state disk are achieved.

A computerized method for detection of format drift and format anomalies is described. A format representation for each data point of a first data sample is extracted. Transformations of each format representation is conducted, resulting in a first plurality of count values (reference) and a second plurality of count values. Each count value identifies a number of occurrences of a transformed format representation within that data sample. Thereafter, a first probability distribution for the first plurality of count values and a second probability distribution for the second plurality of count values are computed. Analytics using the first and probability distributions are conducted to produce a first metric. A format drift is determined based on an evaluation of the first metric to a second metric operating as a threshold metric. Format anomalies are detected based on analytics of hashed format representation and determination of infrequent usage of a particular format representation.

The invention discloses a floating-point number processing method and device, computing equipment and a storage medium, and relates to the technical field of computers. The computing device obtains a first floating-point number, and decodes the first floating-point number to obtain a symbol, an order code and a mantissa. The first floating-point number may include a first symbol field, a bit width indication field, a first order code field, and a first mantissa field. The bit width indication field is used for indicating the bit width D occupied by the first-order code field in the total bit width N of the floating-point number, and when the total bit width is fixed, the bit width of the first-order code field and the bit width of the first mantissa field can dynamically change along with the numerical value indicated by the bit width indication field. In addition, the bit width indication field can also indicate whether the first mantissa field represents the mantissa or the offset of the order code, and by multiplexing the first mantissa field, the influence on the precision of the floating-point number is small, and the coding space of the order code can be increased, so that the dynamic range of the numerical value represented by the floating-point number is expanded, and the requirements of AI training and reasoning can be met.

The embodiment of the invention discloses a game numerical value balance adjustment method and device, equipment and a storage medium, and the method comprises the steps: obtaining game numerical value tables of different types of game numerical values, and carrying out the format conversion of each game numerical value table to obtain a corresponding first data table in a standard data format; associating the first data tables corresponding to the game numerical values of different types according to the game role identifiers to obtain a second data table, wherein the second data table is a linked table associated with the game role identifiers; and comparing the game settlement parameter corresponding to each game role identifier in the second data table with a preset threshold value, and adjusting the corresponding game numerical value according to a comparison result to obtain a game numerical value adjustment result. According to the scheme, the problem that the game numerical value balance adjustment efficiency is low can be solved, balance adjustment of the corresponding game numerical value is carried out based on comparison of the game settlement parameter and the preset threshold value, automation and intelligence of game numerical value balance adjustment are achieved, and the overall efficiency of game numerical value balance adjustment is improved.

The invention relates to the technical field of data processing, in particular to a target neologism determination method and device based on a cohesion degree algorithm and a medium, and the method comprises the following steps: processing a given corpus to obtain a plurality of candidate neologisms, and calculating a plurality of attribute indexes of each candidate neologism, dynamically acquiring an index evaluation threshold value of each attribute index of the candidate new words according to the length of the candidate new words, the frequency adjustment factors corresponding to the candidate new words and a numerical value distribution condition of each attribute index corresponding to a plurality of candidate new words in a given corpus, marking as 0 when any attribute index is smaller than the corresponding index evaluation threshold value, otherwise, marking as 1, and determining that the candidate new words are not the candidate new words. When the product of all the mark values corresponding to the candidate new word is 1, determining the candidate new word as a target new word; according to the method and the device, the multi-dimensional attribute indexes of the candidate new words are obtained, and the index evaluation threshold value of each attribute index of each candidate new word is dynamically adjusted, so that the accuracy and the reliability of screening the target new words are improved.

Techniques for filter and sort versions of software instances executing on computing devices in a network are disclosed. The system receives an instruction to execute an operation on software instances that satisfy numerical criterion and non-numerical criterion. A standardizing function is applied to the numerical criterion to determine a standardized numerical criterion. The system accesses version number values corresponding to software instances on the computing devices in the network. By applying the standardizing function to numerical version numbers of the version number values, the system determines standardized numerical version numbers for the version number values. The system filters the standardized numerical version numbers based on the standardized numerical criterion. The system identifies software instances with (a) associated standardized numerical version numbers that meet the standardized numerical criterion and (b) non-numerical components that meet the non-numerical criterion. The system executes the operation on the software instances that meet the criteria.

The invention discloses a floating-point number format conversion device and method, a storage medium and a program product, and relates to the technical field of floating-point number format conversion. The device comprises a preprocessing unit and a conversion mapping unit; the preprocessing unit is used for receiving input data in a first floating point format and performing bit width compression processing on at least part of digits influencing rounding operation in the input data to generate a corresponding index value, and the bit width of the index value is smaller than the mantissa bit width of the input data; the conversion mapping unit is coupled to the preprocessing unit, a plurality of mapping relations between index value ranges and numerical value representations in a second floating point format are stored in the conversion mapping unit, and the conversion mapping unit is used for searching the corresponding numerical value representations in the second floating point format according to the index values to serve as output data. The invention aims to reduce the number of logic gates and the circuit complexity required by floating-point number format conversion, so that the area of a related chip is reduced, the power consumption is reduced, and the operation delay is shortened.

The invention discloses a data processing method and device, which can reduce service time delay and equipment cost. In the method, a data receiving and transmitting end determines N CB groups corresponding to a first TB according to a first numerical value. The sending end executes LCP with the CB group as the granularity, and the receiving end decodes the CB group with the CB group as the granularity. Wherein the first numerical value is the number X of the CBs contained in one CB group, or the number Y of the CB groups contained in one TB. The N CB groups meet one of the following items: the number of CBs included in each CB group in the N CB groups is a second numerical value; or, the number of the CBs contained in the first M CB groups in the N CB groups is a second numerical value, the number of the CBs contained in the last N-M CB groups is a third numerical value, and M is a positive integer smaller than N; or, the number of the CBs contained in the first N-M CB groups in the N CB groups is the third numerical value, and the number of the CBs contained in the last M CB groups is the second numerical value; or, the number of the CBs contained in each CB group in the N CB groups is a fourth numerical value.

The accuracy of existing machine learning models, software technologies, and computers are improved by using one or more machine learning models to predict a type of data that one or more numerical characters and / or one or more natural language word characters of a document correspond to. For instance, a Question Answering systems can be used to predict that a particular number value corresponds to a date, a billing amount, a page number, or the like.

The invention discloses a database-based fully-encrypted data query method, apparatus and device, and a storage medium, and relates to the technical field of databases, the method comprises the steps of constructing a hash construction result corresponding to to-be-encrypted data by using a preset hash algorithm and a private key, and determining a modulus based on a numerical value type of the to-be-encrypted data; determining a disturbance value based on the hash construction result, the modulus, and the numerical symbol and the numerical type of the to-be-encrypted data, judging whether the current database retrieval scene meets a preset continuous summation condition, and encrypting the to-be-encrypted data based on the judgment result and the disturbance value to obtain encrypted data; and when a ciphertext database corresponding to each piece of encrypted data is retrieved, obtaining corresponding target encrypted data based on the query statement, and decrypting the target encrypted data based on the disturbance value corresponding to the target encrypted data to obtain a query result corresponding to the query statement. Therefore, the high calculation and retrieval overhead of the encrypted data in the database is avoided.